David Moses details 28-employee AI agent rollout using Hermes, MongoDB, and deterministic workflows for real estate ops
Summary
David Moses outlines a goal to equip all 28 employees with dedicated AI agents by November 9, 2026, mirroring user permissions across CRMs like Property Meld, App Folio, and Sync. The infrastructure relies on a single Mac Studio running the Hermes harness, with data aggregated in MongoDB to overcome API limitations in legacy systems. Moses argues for deterministic Python scripts over AI 'skills' for repetitive tasks to ensure reliability and reduce costs. Jack BeVier and Moses debate the trade-offs between AI speed in contract review and human intuition for contextual risk assessment.
Episode sections
- 0:00 – 0:53
Introduction and AI Agent Rollout Goal
Jack BeVier and David Moses discuss an upcoming real estate investor roundtable in Charlotte. David Moses outlines his goal to have every one of his 28 employees equipped with a dedicated AI agent by November 9, 2026, capable of drafting emails and accessing all permissions within their respective CRMs.
- 0:53 – 4:47
Tech Stack and Integration Challenges
David Moses lists the company's software stack, including Property Meld, App Folio, Sync, QuickBooks Online, Z Inspector, Dialpad, Microsoft 365, Zoom, Teams, Fireflies, and Plaud. He notes that many systems lack accessible API audit logs, forcing the company to poll systems to create their own logs for tracking changes.
- 4:47 – 7:45
Data Architecture and Hardware Infrastructure
The company uses a single Mac Studio with 96-128 GB of RAM to host all AI agents via the Hermes harness. Data is stored in a MongoDB NoSQL database for high-speed writes, with backups handled nightly to physical drives and via Backblaze every six hours.
- 7:45 – 11:21
Permissioning and Security Workarounds
David Moses explains that while Microsoft permissions are mirrored for agents, CRMs like App Folio and Sync lack granular API permissioning. To enforce read-only access, the company uses N8N workflows to hard-code allowed commands and endpoints, rejecting unauthorized calls.
- 11:21 – 14:25
Agent Architecture: Hermes vs. OpenClaw
David Moses distinguishes Hermes from OpenClaw, noting that Hermes silos agents so they do not share memory unless explicitly permitted. Agents operate in parallel, and subagents are spawned for validation or complex tasks, ensuring deterministic controls over financial transactions.
- 14:25 – 18:23
Deterministic Automation vs. AI Skills
David Moses argues that AI 'skills' are unreliable soft guardrails. For repetitive, deterministic tasks, he prefers Python scripts or N8N workflows, which enforce strict steps and conditions, increasing reliability and reducing token costs compared to LLM-based approaches.
- 18:23 – 24:03
Operational Efficiency and Contextual Memory
The company uses agents to capture detailed context from work orders, emails, and calls that humans might skip. This creates a comprehensive record in MongoDB, allowing agents to identify patterns and risks (e.g., recurring maintenance issues) that would otherwise be lost.
- 24:03 – 27:15
Employee Adoption and Dependency
Employees have adopted the agents, with 10-15 users having them for over a month. They report enjoying the tools but have become highly dependent, expressing frustration if the internet or agents are unavailable, highlighting a shift in operational reliance on AI.
- 27:15 – 28:34
Data Integrity and Mongo Mirror Architecture
Jack BeVier notes that AI agents maintain higher data integrity than humans by avoiding shortcuts or creative errors. David Moses explains using MongoDB to aggregate data across multiple endpoints, allowing agents to access more information than the native UI of systems like Appfolio provides.
- 28:34 – 32:50
Automated Database Mirroring and Read/Write Logic
David Moses describes an automated process where agents create a schema mirror in MongoDB by scraping endpoints every two minutes. Agents primarily read from this fast, aggregated MongoDB mirror but bypass it to fetch fresh data and sync tokens directly from sources like QuickBooks when writing transactions.
- 32:50 – 37:15
Managing Hallucinations and Model Limitations
Jack BeVier asks how the team handles AI hallucinations. David Moses states they use cheap models for low-stakes tasks where errors are easily ignored, and newer models like GPT-6 do not hallucinate but may fail due to permission issues or throttling during high-volume tasks.
- 37:15 – 40:45
Context Window Management and Hermes Memory
Jack BeVier questions whether agents degrade over long sessions due to context window limits. David Moses explains they rely on Hermes' memory structure and use Discord threads to keep context discrete and relevant, avoiding the performance degradation seen in long, unstructured chat sessions.
- 40:45 – 46:56
AI in Legal Review vs. Human Intuition
David Moses argues AI like Claude is superior to attorneys for contract review due to speed and access to case law. Jack BeVier counters that AI lacks real-world context and human intuition, often missing exposure arising from the history of a business relationship rather than just the document text.
- 46:56 – 52:03
General Counsel Agent and Risk Mitigation
Jack BeVier proposes a 'general counsel agent' that monitors all transactions, emails, and contracts to proactively surface legal exposure. David Moses agrees, noting that honest mistakes accumulate over time and that AI agents writing emails could provide real-time legal guidance to prevent lawsuits.
Claims worth verifying
David Moses's goal is to have every employee at the company have an AI agent by November 9.
Unidentified speaker (UNKNOWN) ·0:53
How to verify: Check company internal communications or press releases from David Moses's real estate management company for the stated rollout deadline of November 9, 2026. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
The company has approximately 28 employees.
David Moses ·1:48
How to verify: Verify employee count via the company's LinkedIn profile, Crunchbase, or SEC filings if publicly traded. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Primary sourceThe company uses Property Meld, App Folio, and Sync for property and community management.
Unidentified speaker (UNKNOWN) ·1:25
How to verify: Check the company's website 'Technology' or 'About' section, or review job postings listing required software experience. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
The company uses QuickBooks Online for corporate accounting.
Unidentified speaker (UNKNOWN) ·5:22
How to verify: Check the company's website or job postings for accounting software requirements; verify via SEC Form 10-K if public. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Primary sourceThe company uses Z Inspector for inspections.
Unidentified speaker (UNKNOWN) ·5:51
How to verify: Check the company's website or job postings for inspection software requirements. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
The company uses Dialpad for phone systems, Zoom for community meetings, and Teams for internal meetings.
Unidentified speaker (UNKNOWN) ·6:14
How to verify: Check the company's website or job postings for communication software requirements. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
The company uses Fireflies and Plaud for meeting recording.
Unidentified speaker (UNKNOWN) ·6:40
How to verify: Check the company's website or job postings for meeting recording software requirements. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
The company uses MongoDB as a NoSQL database for data storage.
Unidentified speaker (UNKNOWN) ·7:06
How to verify: Check the company's technical blog, job postings for data engineers, or GitHub repositories for MongoDB usage. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
All AI agents are hosted on a single Mac Studio with 96-128 GB of RAM.
Jack BeVier ·7:56
How to verify: Check the company's technical blog or job postings for infrastructure specifications. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Backups are performed nightly to physical drives and every six hours via Backblaze.
David Moses ·8:33
How to verify: Check the company's IT security policy or job postings for backup procedures. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
AI agents do not leave off characters or take shortcuts, resulting in better data integrity than human data entry.
Jack BeVier ·27:18
How to verify: Internal audit of data fields in the company's database comparing AI-generated entries to human-generated entries for accuracy and completeness. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
MongoDB allows aggregation of data across multiple endpoints, providing more information than the native UI of Appfolio.
David Moses ·27:53
How to verify: Technical documentation of the company's data architecture or a demonstration of the MongoDB schema versus the Appfolio UI. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
The MongoDB mirror is updated every two minutes by an automation that checks for updates within the last five minutes or hour.
David Moses ·28:59
How to verify: Review of the company's automation scripts or API logs showing the frequency and logic of data pulls from Appfolio and other sources. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Agents bypass MongoDB to read directly from QuickBooks for transaction recording to ensure fresh data and correct sync tokens.
David Moses ·30:16
How to verify: Code review of the agent's workflow logic for transaction recording in QuickBooks. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
GPT-6 does not hallucinate but may fail due to permission issues or throttling.
David Moses ·34:22
How to verify: OpenAI's official documentation or release notes regarding GPT-6 capabilities and known limitations. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
OpenAI throttles long-running tasks during high volume, causing models to stop and require a prompt to continue.
Unidentified speaker (UNKNOWN) ·35:01
How to verify: OpenAI's API usage logs or rate limit error messages during high-volume periods. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Claude is superior to attorneys for contract review because it can instantly pull up relevant case law.
Unidentified speaker (UNKNOWN) ·41:38
How to verify: Comparison of time and accuracy of contract review by a human attorney versus Claude for a specific legal scenario. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
AI lacks real-world context and human intuition, often missing exposure arising from the history of a business relationship.
Jack BeVier ·43:12
How to verify: Case studies or examples where AI missed a legal risk that a human attorney identified based on prior interactions. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
A 'general counsel agent' could monitor all transactions and emails to proactively surface legal exposure.
Unidentified speaker (UNKNOWN) ·46:56
How to verify: Prototype or proof-of-concept of such an agent monitoring a sample dataset for legal risks. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
AI agents writing emails could provide real-time legal guidance to prevent lawsuits.
David Moses ·49:31
How to verify: Pilot program testing AI email review for legal compliance and comparison of lawsuit frequency before and after implementation. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.
Do your own research
- Verify the specific API capabilities and audit log availability for Property Meld, App Folio, and Sync by reviewing their developer documentation or contacting support.
- Check the current pricing and feature set of the Hermes harness and N8N workflows to assess cost implications for a 28-person team.
- Review the security implications of mirroring data in MongoDB and polling systems every two minutes, including potential rate limiting or API costs.
- Assess the reliability of 'cheap models' for low-stakes tasks by testing error rates in image analysis or text generation within the specific real estate context.
- Investigate the legal liability framework for AI-generated contracts and emails, specifically how 'general counsel agents' handle exposure and compliance.
- Compare the token costs of LLM-based automation versus deterministic Python scripts for the specific repetitive tasks identified in the episode.
- Verify the hardware specifications and maintenance requirements for a single Mac Studio handling 28 concurrent AI agents.
- Look for case studies or public reports on employee productivity changes after adopting similar AI agent rollouts in property management firms.
Full transcriptexpand
0:00Jack BeVier: Hey everybody, welcome to another episode of Forked.ai. Uh, this is your host, Jack Bevere, along with my co-host David Moses. David, wonderful to see you, sir.
0:08David Moses: Great to see you too. Looking forward to seeing you in person very shortly here.
0:13Jack BeVier: Yes, yeah. We are convening every four months as we do, a real investor round table. Get about 30 or so very active real estate investors from lots of different markets all over the country. We all get together and kind of share, put put together a soft agenda and share kind of what's going well, what's not going well, the challenges we have, and try to just, you know, kind of use the room as a mastermind to help make our businesses better, come back with actionable things that we can actually improve our businesses. So we got that coming up in Charlotte very shortly next week. And uh yes, I look forward to seeing you there, man. That's gonna be a lot of
0:49Unidentified speaker (UNKNOWN): fun. It will
0:50David Moses: indeed. It will indeed. So
0:53Unidentified speaker (UNKNOWN): so
0:54David Moses: we were talking before we uh before we started this this session here that uh I I I'm very close to reaching my goal.
1:01Unidentified speaker (UNKNOWN): My
1:01David Moses: goal is November ninth to have every every single employee at the company has an AI agent that they can work with. That AI agent is able to drive. draft emails for them. It's able to see everything they can see. So all the files they can see, their agent can see all the permissions they have in the
1:25Unidentified speaker (UNKNOWN): uh
1:26David Moses: various CRMs that they're in and out of. We use property meld, we use App Folio, we use a software called Sync for community management. And so on the community management side and on the property management side, including accounting, even uh even our biz dev, we'll all have an AI agent. We're about five agents away. How
1:45Jack BeVier: many people are we talking about?
1:48David Moses: It is, I think it's twenty eight.
1:51Jack BeVier: Okay.
1:53David Moses: And I'm quite certain that I need to add some more OpenAI tokens, probably I'm guessing three more. But uh because as as we grow, we're we're also I'm I'm building a workflow that tests on it. Well, first, it's kind of frustrating because not all CRM systems have audit logging that is accessible through the API, which it it just sucks in general, because that should never that should never happen. Audit logging in 2026 should really only be it shouldn't even be available through the UI. Just it should be only API. It should be only MCP.
2:29Unidentified speaker (UNKNOWN): It's too much.
2:30David Moses: There's no reason audit logs should be reviewed by a human ever. But anyway, a lot of them just don't have them accessible from the API. And that the the in the underlying issue there being that we can't we we have to essentially create our own logs. So we're polling all the systems so that we can determine what records changed. and most importantly who changed them. So yeah, it's going to give us some really cool information. We'll we'll basically take: look, here's the record, here's what it was before. You know, it'll essentially just,
3:03Unidentified speaker (UNKNOWN): you know,
3:03David Moses: give a summary of any fields that have changed. And if no fields have changed, usually because somebody changed something by accident and changed it back before it pulled the next time, then you know, where that last updated field updates, but no no information has changed. You know, but you have who made the change.
3:24Unidentified speaker (UNKNOWN): And
3:24David Moses: that's important for us because we're we're essentially for adoption for us, I don't it it's it's not going to be based on tokens, right? It's not going to be based on this is how much time and energy and t you know that they spent with their AI agent. Although that's a good way of measuring
3:39Unidentified speaker (UNKNOWN): uh
3:39David Moses: in general. What we really want to measure is how much of the work that you do on a regular basis is actually being done, you know, where the final product is delivered by you without the assistance of AI. And I consider that an accounting person coding an invoice or a property manager sending an email to a tenant or to an owner whose property we manage. a community manager having a conversation with a board member and writing a note about it. All those things should be influenced by what the human is doing and directed by the human, but memorialized by AI. That way, the records are just wildly more useful and the context we need is going gonna to bright is going gonna to be right in the systems that we're already using.
4:27Jack BeVier: All right, let me I have a thousand questions. Just to refresh and contextualize the conversation here, because this is I think a place where a lot of businesses would like to get to and you'd learn a lot getting just getting to this point, right? There's a lot of work, took a lot of work for you to get to the point where you could even do this from a practical perspective. So let's talk about the architecture real quick just to refresh listeners' memories. So you got sync is your HOA uh management system, Appfolio for property management. What's accounting again?
4:59David Moses: We use, well for I I call it app the we we we have ops accounting and corporate accounting. So ops accounting means accounting as an operation for others. We're we're accounting for your your property and how it does, your community, we're doing your books. That's essentially accounting as a service. Those are done directly in those CRMs, like directly in Epfolio, directly in sync. And then our corporate books right now are done in QuickBooks.
5:22Unidentified speaker (UNKNOWN): I
5:22David Moses: would say Yeah, QuickBooks Online. I have no real great defense for it other than it seems to be reasonably reliable in downloading transactions from the actual financial accounts that you know that were where they're housed. I I I see no other reason to choose
5:39Unidentified speaker (UNKNOWN): any
5:39David Moses: one ERP system from another. They're all they're all pretty much.
5:43Jack BeVier: And did any other any other sources of data, any other softwares that you're using as a
5:49David Moses: Yes.
5:51Unidentified speaker (UNKNOWN): Oh
5:51Jack BeVier: yeah, property meld, sorry, for maintenance.
5:53David Moses: We currently use Z Inspector or inspections. We probably won't unless they uh very quickly get an API. We're we're building little
6:02Unidentified speaker (UNKNOWN): our
6:02David Moses: own little mini inspections
6:04Unidentified speaker (UNKNOWN): in
6:04David Moses: hopes eventually we'll either replace them or we don't really want to replace them. We'd rather just use them. But
6:09Unidentified speaker (UNKNOWN): if
6:09David Moses: they don't have an API.
6:11Unidentified speaker (UNKNOWN): If they're an
6:11Jack BeVier: API though. Yeah.
6:12David Moses: Yeah. It's kind of useless to us.
6:14Unidentified speaker (UNKNOWN): So
6:14David Moses: we use you know we use the inspector. Obviously we have our phone system through dial pad. We've got our emails Files through Microsoft and files files uh file storage through Microsoft as well. And our uh meetings are generally done if they're on the community side in Zoom because communities kind of demand it. They just I don't know, they just trust Zoom and uh and then we use Teams for most other other types of meetings.
6:40Unidentified speaker (UNKNOWN): And
6:40David Moses: we use a software called Fireflies
6:42Unidentified speaker (UNKNOWN): and
6:42David Moses: a software called Plaud. So if Fireflies does Because you're in a web meeting, it'll jump in no matter what the platform is, whether it's GoMeet or it doesn't matter. It doesn't matter the platform, it'll it'll jump in, record, take notes. And then and Plaud is like for live recording. Okay. I'm like I'm I'm at a meeting, I'm in a meeting face to face. I use Plaud for recording and a couple other people at at the office have that.
7:06Unidentified speaker (UNKNOWN): And
7:06David Moses: then their phones, they they'll go in and record, and then they just the idea is that they'll give those to their agents, and then we have a Mongo. Mongo. Mongo is basically just our repository for everything. It's a No SQL database. It has very, very quick write capabilities. And it I think in our businesses, it's very common to write a whole lot of data that most of which you're never gonna need to retrieve at any particular time for any particular reason, but it's just y you just need to need to accumulate it. So Mongo gives us very quick write access.
7:40Unidentified speaker (UNKNOWN): And
7:40David Moses: that's that. That's basically our stack our tech stack. And are
7:44Unidentified speaker (UNKNOWN): you those are
7:45David Moses: the
7:45Unidentified speaker (UNKNOWN): and
7:45Jack BeVier: you're using Hermes as an agent? Cloud
7:49Unidentified speaker (UNKNOWN): like
7:49Jack BeVier: you just you go in Hermes and you do you the do the cloud agents?
7:53David Moses: Uh our it's all hosted on a Mac studio.
7:56Jack BeVier: All hosted on a Mac studio?
7:59David Moses: One singular Mac Studio hosts all of our agents. It is it has been a challenge to get to enforce the agents to save things to my to to OneDrive and SharePoint instead of saving it to the local drive. Because like all of a sudden it'll be like 90 gigs was added in a day from some agent doing some long running thing. And uh usually it's me who's created the issue. And then I've got to say, hey, migrate all this stuff off of here. We don't have unlimited storage on this little on this little Macs. Mac Studio. But it's a it's a robust MacD. It's 96 gigs of RAM, or maybe it's 128 gigs of RAM. It's a powerful machine and it yeah, it hosts everything. We back it up to a physical drive every night, and then we also back it up using Backblaze. And I think that's every six hours.
8:44Unidentified speaker (UNKNOWN): And
8:44Unidentified speaker (speaker_2): that's that.
8:45Jack BeVier: So you're the agents that are running on this Mac Studio, did you create seven? separate virtual mach like how how did how you have multiple agents running on one Mac Studio?
8:56Unidentified speaker (UNKNOWN): So it's one
8:56David Moses: harness, multiple agents. So it it's only one Hermes instance.
9:00Unidentified speaker (UNKNOWN): Okay.
9:01Jack BeVier: It's
9:01Unidentified speaker (UNKNOWN): only one
9:01David Moses: Hermes instance. They just it's just all of our we have our our essentially our call it our uh our chief technology. officer agent who basically builds the agents, decides permissioning is actually kind of easy because
9:21Unidentified speaker (UNKNOWN): it
9:21David Moses: just mirrors whatever Microsoft's permissions are. If they can see the folder,
9:26Unidentified speaker (UNKNOWN): they
9:26David Moses: can see the email, they can see the that's all that's all there. And then on the on the on the side of the sync and app Appfolio, folio,
9:36Unidentified speaker (UNKNOWN): the
9:36David Moses: API doesn't let you do everything. You can't pay bills from App Folio or from Sync. But through the API, there's really no way of permissioning one agent's work, like hard permissioning it, so that they literally can't make a call in sync. So we have workarounds for that, which is we'll will either have an end-in-to-end end it end workflow that basically basic like if somebody would only have read access to a particular system. So let's say I I've got to have read access to QuickBooks because I've got to be able to check to see if this if this was invoiced. I have to be able to pull out the invoice. I have to be able to, you know, whatever I have to do, but I I but I only want read access. You could actually create a, you know, for QuickBooks, you can create a key that's like this is only a read access API key. So you don't have to worry about ever any anybody making a a call that
10:25Unidentified speaker (UNKNOWN): that
10:25David Moses: they're not allowed to make. That doesn't hold true in sync or app folio. So what we have to do is we create these workflows in N8N where we hard code the call. So it's either this is how you make a call, we have a separate API key that we generate. And this little very small, very easy N8N workflow, all it does is receives the the call formatted exactly the way it's supposed to be formatted. for the app folio API,
10:53Unidentified speaker (UNKNOWN): but
10:53David Moses: then it hard codes the command. So it's either, you know, if they have
10:57Unidentified speaker (UNKNOWN): read
10:58David Moses: only if they only are getting read access to a particular endpoint, or if they only have access to certain endpoints, then it will allow them to make a call using any of those endpoints or using any of those commands. But it's hard hard-coded coded in the end N at end N workflow that it will reject it if it's not an accepted command or endpoint.
11:18Unidentified speaker (UNKNOWN): So we
11:18David Moses: have to create our own permissioning behind that.
11:21Jack BeVier: Is that done at the So you have one Hermes agent that has access to all of these systems, but then the subagent you manage permissioning
11:32Unidentified speaker (UNKNOWN): on
11:32Jack BeVier: the subagent agent level by employee and their role?
11:37David Moses: Yeah. So okay. So just just a little difference between from what I understand, and I may be wrong on this, but OpenClaw versus Hermes. So in OpenClaw, it's this shared agent ecosystem where they all share the the the memory, they all share the skills. It's all a big C. Whereas with Hermes, it's it's different. Each agent has their own little pond, right? That that they that they live in. And they can only talk to each other if we give them permission to talk to each other. They're fairly well siloed. Their memories are, you know, one one agent can't access another agent's memories unless
12:16Unidentified speaker (UNKNOWN): we
12:16Jack BeVier: sorry to be pedantic, but and you mean subagent right now, right? Because there's only one Hermes agent.
12:22Unidentified speaker (UNKNOWN): Aaron Powell No,
12:22David Moses: Hermes is the harness. So Hermes is the harness. So all the all the agents. are spawned, right? Like they're
12:30Unidentified speaker (UNKNOWN): they're they're
12:31David Moses: operating in parallel. They're not
12:32Jack BeVier: they're operating in parallel. Okay, okay.
12:37David Moses: So so then they can spawn each agent that we can actually interact with. So there's subagents that are temporary subagents. You create two subagents, sub
12:46Unidentified speaker (UNKNOWN): agents. you're gonna run a
12:47David Moses: away. Yeah. Th Those are things that you wind up doing like when you're using things like Claude and OpenAI, like and and Grok for sure, they they'll spawn the sub agent. They don't tell you how it's doing it, right? Like it's it's it's getting it right. That's all you care about. But
13:05Unidentified speaker (UNKNOWN): in
13:06David Moses: some instances it's spawning multiple sub agents as it sees that it's necessary in order to make sure or validate what it's what it's telling you. And of course that's still happening. with our agents, but with our agents, because we want, you know, we're we've got live data, we've got, you
13:21Unidentified speaker (UNKNOWN): know,
13:21David Moses: these are financial transactions in many cases. They're, you know, these are this important data that affects people's lives. So we typically when we're doing uh anything, there is there is at least a validator agent by rule. You're gonna enter this tr you're gonna enter this uh you know d this transaction into QuickBooks. You have to have a validator agent that looks at the call you're about to make, looks at the transaction and says, I agree with uh with how the first agent determines to do it. But that's all sub agents kind of behind the scenes. We don't talk to those sub agents. We talk directly to the agents.
13:54Jack BeVier: Is the validator agent, for example, or like the corporate pr like the the corporate guidelines policies and procedures, is that memorialized?
14:03Unidentified speaker (UNKNOWN): in
14:04Jack BeVier: an age in a in a in a you know a permanent agent state that acts as a you know a manager, so to speak, to make sure that one of the employees' agents doesn't perform a task which is outside of policy? Or are you relying upon the direction of the employee to be within guidelines, within guardrails?
14:25David Moses: So there is obviously reliance on the employees, but there's also the every single thing you ask your agent to do is memorialized forever. Right. So if you ask your agent to go look in David's emails for anything that could be, you know, used to blackmail him, like,
14:43Unidentified speaker (UNKNOWN): you
14:43David Moses: know, it it knows. It it it remembers that she said that and in and but there it, you know, and and then we run like you know, hey, look through all of the agent's activity and tell me if you see anything that's a security risk or that's uh potentially a you know a legal risk. And we, you know, but but no, like we don't have a master agent that's and and that's probably something we need to do. Like we don't have a master agent that's literally monitoring on a regular basis what each individual is discussing with their agent, right? What we also what we generally do in terms of enforcing policies is that they use skills. And I'm learning that skills are not a reliable way to automate anything. Skills are soft guardrails guidelines at best in the Hermes. world, they're probably not much better in any other system, if they are better at all. So what I'll say is if it's something that needs that is repetitive, that needs a reliable
15:53process, it should be, you know, an automation software like NADN made in should be engaged. Or you can do it as a Python script. You can do it, you can create scripts around it, but those scripts need to be largely deterministic. And so think of it this way: it's like a skill is basically a recipe for how to make soup. Right. And it tells you all the ingredients, it tells you what temperature to use. It tells you uh in what order to put those ingredients and how long to wait between. But
16:25Unidentified speaker (UNKNOWN): you're
16:26David Moses: a human being and if you think it, you know, last time you made it it wasn't spicy enough, or if you think whatever, you can just mess with it however you want to mess with
16:35Unidentified speaker (UNKNOWN): it.
16:36David Moses: Whereas
16:37Unidentified speaker (UNKNOWN): a
16:37David Moses: Python script or an N8N workflow, you're gonna go into the kitchen and you know, the all the ingredients are pre pre-measured measured for you sitting where they're supposed to be, and you're assembling them and and and you don't have there's it's no you don't have like extra garlic sitting around right it it's just this is these are the ingredients you get and this is the
16:58Unidentified speaker (UNKNOWN): the
16:58David Moses: steps that are enforced the oven literally won't work until you do this it won't the the oven only has one temperature and it's the temperature you're supposed to use so so that's that's the kind of thing when you when you make it into a Python script or an N8N workflow that's
17:14Unidentified speaker (UNKNOWN): the kind
17:15David Moses: of thing that gets enforced. It just increases reliability considerably for us. And then we can and I just I like the visual nature. There's nothing inherently better about a Python script or about an N8N workflow. They they essentially perfor they do the same thing, but one is visually
17:34Unidentified speaker (UNKNOWN): one
17:34David Moses: one you can look at visually and see it firing. And I like that.
17:39Jack BeVier: Yeah. And there's no reason to leave something that is deterministic, that you want to be deterministic, you want to be an if-then statement. There's no reason to leave it to chance. Even if the agent would get it right ninety-nine out of a hundred times, just write it in code and it yeah, it never never gets right.
17:57David Moses: It's
17:57Unidentified speaker (UNKNOWN): a
17:57Jack BeVier: hundred
17:57Unidentified speaker (UNKNOWN): people. And you're not
17:58David Moses: And you're not wasting tokens on something
18:00Unidentified speaker (UNKNOWN): that's completely.
18:02Jack BeVier: The script's much cheaper to run than the tokens to figure out the result problem of the problem. So how are you are the agents the super interesting? Are you are the agents per person? Like is it like each person's got an assistant and that assistant has a name now? And they're essentially assistant, whatever the role was, to just take stuff off of that person's plate? Is that the
18:23David Moses: So Automations are generally automatically do this for me every time? Those are automations that kind of they go a level up and somebody else builds them. But
18:36Jack BeVier: somebody
18:36Unidentified speaker (UNKNOWN): else builds them and
18:37Jack BeVier: then where where they were where they compiled.
18:39Unidentified speaker (UNKNOWN): Yeah.
18:40David Moses: Say that again.
18:41Jack BeVier: All of the automations and NA to end flows, where are those running from? Where those the are
18:46Unidentified speaker (UNKNOWN): they're
18:46David Moses: running on an EC2, like an Amazon EC2 uh server? to a cloud
18:50Unidentified speaker (UNKNOWN): server.
18:51Jack BeVier: Is that just a cloud server? Okay.
18:52Unidentified speaker (UNKNOWN): Yeah
18:52David Moses: it's a cloud server.
18:53Unidentified speaker (UNKNOWN): So
18:54David Moses: it's yeah it's nothing you know it's it's it it's good enough.
18:59Unidentified speaker (speaker_3): Mm-hmm.
19:00David Moses: And and f Fable, Opus, even even Soul and Astra uh from from OpenAI can manage it really well. So that's good. And and we have but but the way the way we're using it is really it's more along the lines of you were going to go enter this transaction. Don't enter that transaction. You are going to go update this tenant's record. Don't do that. Right. You're going to tell your agent what you want to do. Your agent is going to go scour all of our systems that it can access. The thing like if I ask you, it's if I said your only job today like, I'm gonna I'm gonna pay you your salary. You're like, you're not gonna get there's there's no I'm not gonna dock your money, but but all I want you to do, well let's say for the next two hours, is I want you to record this one what happened on this one work order. That's all I want you to do. Right. And you need to fill the time. Right? Some people are gonna twiddle their thumbs and fuck around. But some people are are are gonna overdo it, right? They're gonna they're gonna go they're gonna look at every single picture. They're gonna like, you know, kind of write a little note about every single picture, and they can look at all the pictures together and see, okay, did that tell the whole story? Is there something missing? They're gonna look at what the what the tenant called in originally, and then they're gonna look at all the communication between the
20:19tenant and the technician or the tenant and the property manager along the way. And then they're gonna look at the closing notes and look at the closing pictures, and then when they're all done. with that, they're gonna, they're gonna put this incredibly detailed set of notes into some system, right? And
20:35Unidentified speaker (UNKNOWN): I'm
20:36David Moses: so to some degree I'm oversimplifying this. But it would be it's extreme overkill, and no one would ever do that who worked for you un unless you look literally force them to do that, because it just feels inefficient and wrong and stupid, right? Even though one out of every eleven times they do that, maybe one out of three times they do that, depending on the technician, they're gonna find something really fucked up. They're gonna find something that just shouldn't have been done that way. Or they're gonna find something missing, or they're just gonna miss something that some other stakeholder is going gonna to call and ask about six months. from now,
21:15Jack BeVier: right? Yeah.
21:16David Moses: And hold their feet to the fire on it. And so
21:19Unidentified speaker (UNKNOWN): and
21:19David Moses: so what I tell them is like just there's no point for you to go in and put in a closing note. There's no point for you to go in after having this conversation with this tenant and go and create this work order with this ticket. It's it's silly. Your agent is gonna going to go grab all the context, it's gonna going to look at oh, how many times has this person called before? about the same issue? How many times did the previous tenant call at this unit about the same issue? Are there other tenants? If it's a multifamily, are there on a other tenants in the same building? Where are those units relative to this unit? All of these things that if you build a workflow around it, all these things are going to get taken care of. But we can't build, we just can't build automations and workflows around every single thing that we do. So
22:03Unidentified speaker (UNKNOWN): a
22:03David Moses: and and it was we were beating our heads against the wall because
22:07Unidentified speaker (UNKNOWN): a
22:08David Moses: lot of this stuff is really common sense. And the agents will learn from our people's experience much faster if we just give them an agent. Talk to your agent. Tell your agent what you're trying to record. What what is what is the solution? What what is the what is the outcome that you're looking for? And the agent is going to go and look at all of the previous times where this thing was recorded, which this is how it'll build on itself over time. Because as though as the previous work gets better and better, those are the good examples that the that the agents can look at, look back at when they're trying to do these tasks in the future. So don't do it. Just have your agent do it. Your agent's gonna go to all the systems, grab all the context it needs, grab all the phone call transcripts, grab all the emails, grab all everything that's that's not where it's actually supposed to be, right? Uh, because I had a conversation with this tenant about the fact, or somebody in the office had a conversation about with the tenant about the fact that they're 11 days behind on their rent, and they mentioned something about this work order, and the person who's talking, they just wanna they just want the money, right? Just just pay us, all right? They're not they don't care about the the jiggly door handle, or they don't care about
23:18Unidentified speaker (UNKNOWN): the
23:18David Moses: fact that they said that the lawn guy was cutting the lawn and a stone flew up and and hit the basement window and cracked it. Like the person collecting money doesn't care about that. But then that contact that context gets pulled in and that agent tells the property manager, Hey, wait a second, you know, before we record this, uh, you know, this this three months later, the tenant finally calls in the work order because, you know now the crack is is is now like some animals come in and whatever. Right. And now we go like now the context can be hey go back to the lawn guy and be like, hey dude, you were here, you cracked the window, whatever, whatever. And and it you know
23:57Jack BeVier: Mystery solved. Yeah.
23:59David Moses: Exactly. Like it's not and
24:01Unidentified speaker (UNKNOWN): without
24:01Jack BeVier: digging for an hour. Yeah.
24:03David Moses: It's right. Without without trying to solve like I don't understand how did this, you know, situation go from how how did it get to this point? And and I think that the more and more tools we give them and the smarter the agents are, the the more fun their jobs seem to be. The people I mean, we have twenty-something 20 something people that already have these agents. Probably only maybe 10 to 15 of them have had it for more than a month and a half. And they love it. They absolutely love it. And it's like it they have to retrain themselves to rely on the agent. And then they get super pissed off if the agent is down for a minute or if, you know,
24:45Unidentified speaker (speaker_2): it's
24:45David Moses: like
24:45Jack BeVier: the internet's out.
24:48Unidentified speaker (UNKNOWN): What
24:48Jack BeVier: are we going gonna to do?
24:49Unidentified speaker (UNKNOWN): The
24:50David Moses: whole
24:50Unidentified speaker (speaker_2): world is fault. Like what's going on?
24:52Jack BeVier: Everyone go home. The internet's out. Everyone go home. No, no work was done before the internet. Go ahead.
24:57Unidentified speaker (speaker_2): If I if I if I gave you a choice between losing power and losing internet
25:03Jack BeVier: for
25:03Unidentified speaker (UNKNOWN): for
25:03Unidentified speaker (speaker_2): a let's say a day, you'd be
25:05Unidentified speaker (UNKNOWN): like
25:06Jack BeVier: 10
25:06Unidentified speaker (UNKNOWN): times out of ten minutes. Yeah, turn
25:07Jack BeVier: the lights off. As long
25:08Unidentified speaker (UNKNOWN): as my phone connects, I'm good.
25:10Unidentified speaker (speaker_2): I got batteries. We're good.
25:12Jack BeVier: Yeah. That's hilarious. So are the agents, are these agents connecting to each of those each of those softwares, each of those CRM database, or are they directly? Or are they connecting to Mongo, which is pushing and pulling the data to each of these derivative systems of record, I'll say?
25:33Unidentified speaker (UNKNOWN): Aaron Powell
25:34David Moses: So sometimes and and this is it's kind of tricky because the a this is where long running things with the right skills and the right memori you know, the kind of the right recipes, this is where the agents get really smart and start to use Mongo for because Mongo mirrors everything in all of our systems.
25:51Unidentified speaker (UNKNOWN): That's
25:51Jack BeVier: what I was
25:51Unidentified speaker (UNKNOWN): gonna ask.
25:56Jack BeVier: Is it it's a reflection of all of those databases? And are they all keyed together? so that the the the unique keys or the common keys between each of those databases are mapped?
26:10David Moses: So
26:10Jack BeVier: does the you know, does the QuickBooks GL code does does an agent know that the QuickBooks GL code for this asset is that property in Appfolio?
26:20David Moses: Um
26:20Unidentified speaker (UNKNOWN): or
26:21Jack BeVier: is it just be like, oh look, the address is pretty much the same and it figures it out?
26:24David Moses: It figures it
26:24Unidentified speaker (UNKNOWN): out. It
26:25Jack BeVier: figures it out.
26:26David Moses: We we get it. like
26:27Unidentified speaker (UNKNOWN): that. You didn't
26:28Jack BeVier: do the work of doing that. The agent's doing a good enough job that you just Okay.
26:32Unidentified speaker (UNKNOWN): Yeah,
26:33David Moses: I'm trying to think of it. I was just trying to think of a situation where we actually do hard code it because the some of the early stuff we kind of had to. But now, you know, we just got smart about how we memorialize things and how we, you know, like
26:46Unidentified speaker (UNKNOWN): uh
26:47David Moses: an example is if it's a work order,
26:49Unidentified speaker (UNKNOWN): we have
26:50David Moses: a separate a maintenance company, right, from our from our management company. And
26:55Unidentified speaker (UNKNOWN): when
26:55David Moses: a work order gets created in property meld, a subcustomer gets created in QuickBooks. And the name of the subcustomer, the subcustomer is a subcustomer of the unit, which is a subcustomer of the building, which is a subcustomer of the owner. And so the sub the name of that, the the the the furthest down the the chain subcustomer is the work order number. So it's like
27:18Jack BeVier: Yeah. Yeah, as the agent's doing more of that work, right? Like the the agent doesn't leave off the N dot in a n you know north. You know, it it just it just takes that and it's the same. And then so just because it it never is creative or lazy or you know takes a shortcut, it just you know types the it just types the the letters, the exact same. And so the data actually the data integrity created by the agents is actually much better than the data.
27:53David Moses: Or more more often it just forgets that this is where you get that information, not over here. And that's what I was talking about with Mongo. So like Mongo lets them aggregate data across multiple endpoints and actually get more information than the data than they can actually get out of the database than even a human can get out of the database. You know, the actual UI going into Appfolio, there are there are things that relate to each other that aren't intuitive in the way Appfolio designs their database and the way they even design the UI, that it's super because in Mongo,
28:26Unidentified speaker (UNKNOWN): all of
28:26David Moses: these things, you know, taken in aggregate, can get them the data that they need.
28:34Jack BeVier: So do you have a schema in Mongo per software that it's connecting to?
28:39Unidentified speaker (UNKNOWN): It
28:39David Moses: it's a mirror, right? So every single endpoint, every single field, however frequently it gets so some of them have a last updated.
28:49Unidentified speaker (UNKNOWN): So you just
28:49Jack BeVier: told you just told the agent to, hey, create a mirror in Mongo of this database. And it just suck it's like, okay, give me all the endpoints, create a schema, map, start writing.
28:58Unidentified speaker (UNKNOWN): Yep.
28:59David Moses: And it's an it it's an automation. It just goes just literally flips through every single endpoint, looks for field like if it's a if it's a an endpoint that has a last updated date, it just looks ha has it last updated in the last five minutes and there's it runs every two minutes. So it's like there is a little bit of overlap there.
29:17Unidentified speaker (UNKNOWN): And
29:17David Moses: if it's if it's updated in the last five minutes or an hour, depending on the database and what the API call limits are and whatever, it pulls the data in, puts it into Mongo exactly the way it is in the system, is if you were to make that API.
29:32Jack BeVier: So the agents are mostly well, I mean you don't know, but the mate ma the agents are analyzing the Mongo data primarily because it's all in one spot. Yeah,
29:43David Moses: it's fast.
29:43Unidentified speaker (UNKNOWN): And it
29:44Jack BeVier: and pattern recognition is like, you know, really, really great. So it does that really, really well. And then you just decide when you tell it to write to the other ones to keep everything to keep everything updated.
29:56David Moses: So the when it r it's gonna read data typically, typically it's gonna read data from Mongo. Sometimes it goes to the actual record because if the agent reasons that it needs up to the minute, like if it's gonna record, if it knows that the end game is to record a transaction in QuickBooks, it doesn't bother going to Mongo to pull the information about that transaction. Because in QuickBooks it needs a sync token. And so by grabbing that transaction from QuickBooks, it knows it's the freshest data and it knows the sync token is correct. So
30:35Unidentified speaker (UNKNOWN): it's
30:35David Moses: not it doesn't bother going into Mongo, right? But if it's hey, give me a report, or if it's giving a report, it knows it's read-only. I'm gonna go to Mongo. It's faster and it has it has more information. Because it the automations that like we fill in the gaps when we create the mirror database. we get smart we get smarter about filling in the gaps where Appfolio and Sync doesn't give us information that we actually want. So we have now we're creating this logging, you know, that just tells all you know tells It us in every database what's changed. But there are some there were some fields and some,
31:16Unidentified speaker (UNKNOWN): I
31:16David Moses: guess, some information that was so key to what we were doing that we we needed to s we needed to start accumulating right within that record or that that yeah, right within that record itself. We needed to add more information that wasn't there before. Right. So add information that Appfolio wasn't giving us. And we
31:35Unidentified speaker (UNKNOWN): potentially
31:36Jack BeVier: you're never writing to Mongo. Mongo is always just pulling in and a r and a mirror. It's just always a mirror, but whenever you write, you write to the
31:45David Moses: Correct. It it yes. We uh th there's no it's not
31:48Unidentified speaker (UNKNOWN): two way.
31:49David Moses: Yeah. It's a because the the connection from Mongo to the to the database is not two way. Uh that's not to say we don't still write to Mongo directly for other things. We'll we'll have things that just don't live anywhere else that are in Mongo. Like we we'd
32:01Unidentified speaker (UNKNOWN): write write. Are
32:02Jack BeVier: things like property mailed and appfolio connected so that when a maintenance ticket gets created or dispatched and resolved in property mailed, that Appfolio is automatically updated? Or
32:13David Moses: that's already connected between the
32:15Unidentified speaker (UNKNOWN): property connected.
32:16David Moses: But we kind of ignore that connection. We don't really it's not helpful for us because we we don't
32:22Unidentified speaker (UNKNOWN): the
32:22David Moses: reason we don't like if if we were satisfied with what was what Appfolio's capability capabilities were, and I I say were because I I don't know what it is today. Um, but if we had been happy with how it deals with work orders, we would have just used it. We wouldn't have bothered using a third party. So there's no point to extract data from App Folio about the work orders, but there is a connection there is a connection that's automatically made there.
32:50Unidentified speaker (UNKNOWN): Okay,
32:50Jack BeVier: so then I gotta ask when
32:53Unidentified speaker (UNKNOWN): because
32:53Jack BeVier: I I gotta think that the agents hallucinate sometimes, when they hallucinate, what do you do about it? Do you bump pump up the mo pump up the model strength and hope that they into until they don't hallucinate and just throw money, which is basically just like throw money at it until
33:07Unidentified speaker (UNKNOWN): they don't
33:07Jack BeVier: hallucinate? Or
33:09Unidentified speaker (UNKNOWN): are
33:10Jack BeVier: you giving them additional context? If so, where is that context stored and how is it shared with everyone? every other agent so that the next agent doesn't do the same hallucination a month later? How do you manage hallucinations?
33:22David Moses: I don't. And I'll I'll I'll kind of explain why. I if it's important enough that like the only things I use kind of cheap models on are things that are within a work. workflow that's running where the hallucination is not going to be where it's going to be a subjective hallucination. Like
33:47Unidentified speaker (UNKNOWN): it
33:47David Moses: may look at a picture, right? May process a picture and it may say, this is out of square, or this is, you know, there's there's I see rust accumulation here. You know, those are things that like that's so there are so many ways that it can like it's
34:05Unidentified speaker (UNKNOWN): not
34:06David Moses: likely to say there's rust there unless somebody looks at
34:08Unidentified speaker (UNKNOWN): the range. Yeah, you can just ignore
34:09Jack BeVier: it. It gives you some, you know, it shortens your analysis cycle
34:12Unidentified speaker (UNKNOWN): time,
34:12Jack BeVier: makes it lighter.
34:14David Moses: It raises the red flag, we look at it. Eh, yeah, I agree.
34:16Unidentified speaker (UNKNOWN): No, I
34:17David Moses: don't agree. That kind of thing.
34:18Jack BeVier: If it's really annoying, you might go back and change the prompting, but
34:21David Moses: that's it.
34:21Unidentified speaker (UNKNOWN): Usually Exactly. it's
34:21Jack BeVier: fine.
34:22David Moses: Yeah. I I we don't we don't account for him because everybody's agent is on. GPT six Astra or GPT six soul. And they just don't hallucinate. They they f they they make a lot of mistakes, but they're not hallucinations. They're I don't have permission to do this. Yes, you have permission to do this. I, you know, it it will not follow the steps. It will not complete the assignment. It'll just like say, Okay, I did the first part of it. What do I do now? And you're like, I why did you just do the first part of it?
34:52Unidentified speaker (UNKNOWN): Is
34:52Jack BeVier: that
34:52Unidentified speaker (UNKNOWN): because the
34:53Jack BeVier: skills files? Is it is that is that
34:54Unidentified speaker (UNKNOWN): resolved through more
34:56Jack BeVier: robust skills file files?
34:58David Moses: I I suspect it's there's two things going on. Number one,
35:01Unidentified speaker (UNKNOWN): the
35:01David Moses: models only have so they they just there's only so much bandwidth out there. There's there's only so there's only so much and and I think in high volume times, they just OpenAI just throttles the long running tasks. and you're gonna get this kind of like it just goes to you. And then you just prompt it and it just keeps going. It picks up where it left off. And then the other thing is that we have inconsistent rules. So it's like when we're building a workflow, and this is really a challenge for us. When we're building a workflow and you're not using Fable, because I say that because Fable's so good that it mo it one one-shots shots most, you know most things that you understand what you want really well, it will one shot.
35:43Jack BeVier: Yeah, I agree.
35:45Unidentified speaker (UNKNOWN): But
35:45David Moses: but when but when you're working in like you're building a workflow or tweaking a workflow using GPT, it will it'll build something and it'll keep going on tangents that you know keep going down a road that's the wrong road, and you don't find that out.
36:03Unidentified speaker (UNKNOWN): in
36:03David Moses: a super long running task until it's done. So we have checkpoints, we have milestones that the when Carl, who's actually my brother, or when Jay, or when somebody else in my office is working on a workflow, it'll stop at milestones deliberately to say to check in with the human and say, Am I am I doing this right? And then we'll okay. Yeah, you're doing it right. So but that's that spills over, that programming spills over into how that main agent builds the other agents that it builds for the for the individuals. And so some of these skills, they wind up stopping at these milestones that really didn't need to stop. The skill is clear, it's very easy, it's not going to mess it up. Just just just go.
36:51Jack BeVier: Hmm.
36:52David Moses: And
36:53Jack BeVier: what is the what is the con I'm going to use the the Claude example. If you have a session in Claude and you've been working in that session for a long time, that entire session is context and is processed for each subsequent request that you have. And so it gets one expensive to two uh worse at its job. as you reach a certain inflection point in the size of that session, the size of that context window. How does Hermes treat that? Are these agents, everything that they've ever done, a context window? Or does it analyze context discreetly and then just call the relevant skills? How how how does it? I would think that you know, an employee, a human doesn't get worse at their job. over time. They get better at their job over time. What's that?
37:44David Moses: An agent, that's not true. Yeah,
37:46Jack BeVier: an agent that's not true. They get worse at their job over time. The more and more instructions you give them, the more and more background they give them, it gets it's laborious for them to process so much, right? So like and they actually get worse at it. How do you r resolve that or or deal with that?
38:01David Moses: So we obviously we're really heavily reliant on Hermes memory structure, which is good and gets better. How it
38:08Jack BeVier: Because Hermes, it's Hermes as a harness is building skills and context based off of those in an in what we think is a an efficient way such that you as the user don't even have to worry about it. It's just happening.
38:22David Moses: Yep. That's exactly right. And then there's the other part of it, which is what the user can do. And so we have two things. We use Telegram
38:31Unidentified speaker (UNKNOWN): and
38:31David Moses: we use Discord. And we've talked about this a little bit before, but Telegram is like, hey, go add this to my calendar. Hey, what's my day look like tomorrow? Hey, can you go summarize the last two phone calls I had with this guy? I've got to call him. Those are the types of things that we would do in Telegram. They're like one one-off off things. But if you're if you've got a job that you do repeatedly, right? Like I'm reviewing statements, I'm analyzing complaints, I'm cheese, uh, whatever, whatever it happens. I'm recording closings, right? Whatever it is that the agent is doing in Discord, you have a thread for it, right? You'll have a channel for the agent or a channel. chan you might have a channel that's just devoted to one overall overriding topic, and then you'll have channels
39:20Unidentified speaker (UNKNOWN): under
39:21David Moses: and you'll have uh sorry, you'll have threads underneath it. And so that helps because the because Hermes understands and has, and we've actually built this into its structure, is that the primary context for whatever it is you're working on is gonna going be to in the same thread.
39:39Unidentified speaker (UNKNOWN): Right.
39:39David Moses: How you've done it before, it's going to be in the same thread. So if you're unsure, you're going to go back and review that thread. And sometimes you have to tell it,
39:47Unidentified speaker (UNKNOWN): go
39:47David Moses: back and review this thread. We've already discussed this.
39:50Unidentified speaker (UNKNOWN): And
39:50David Moses: it and it will do that. But it's a very small amount of data to review for the agent because it's all just this one topic, kind of similar to how you create conversations in Cloud. Claude, where you've got like but it's they're more it's it's more per it has some more permanence to it, right? Like a conversation is just, hey, something I'm working on right now. Whereas this these would be like a permanent conversation where it's not just the conversation I'm gonna have about this one thing, but all of the conversations about all of the things related to that one thing. And I kinda I kinda wish Claude would do the same thing. It's kinda kind of it's kinda kind of silly that they don't, but
40:26Unidentified speaker (UNKNOWN): they
40:27David Moses: have projects and then they have conversations, but you
40:30Unidentified speaker (UNKNOWN): know,
40:31David Moses: it's uh projects
40:32Unidentified speaker (UNKNOWN): are a little more
40:33David Moses: robust.
40:33Jack BeVier: Yeah, the the that that harness is not intended to be an assistant. You know, it's very gatekeeped. It's meant to write code on big projects and one-shot like, you know, software.
40:45David Moses: Yep. Yeah. I I the legal profession, by the way, is like like i if I were if I were going to be in any position right now, any profession right now, where I would believe that I'm in trouble, I would probably say the legal profession has got to be right up there because of the dollars involved. I mean, you have. So how many times do you call and have a conversation with the attorney about a contract that you want to tweak or a whatever it is, right? A demand you got or a lawsuit or you whatever doesn't matter what it is. Claude will give you such a better answer. And s and and this is no offense, but your attorney's not going to put eleven hours into figuring out.
41:38Unidentified speaker (UNKNOWN): Yeah.
41:38Jack BeVier: Well, you don't
41:39Unidentified speaker (UNKNOWN): want 'em
41:39Jack BeVier: to. Yeah.
41:40David Moses: You don't want 'em.
41:41Unidentified speaker (speaker_2): It's too expensive. But
41:42Unidentified speaker (UNKNOWN): if
41:42David Moses: you go into Claude and you just say, hey, I want to tweak this contract, it's like, well, you're missing this clause that because of this law, or da da da, and I want to add this. And you're like, well, can you just like from a practical sense, I really want to understand why this clause is why why I need this clause written this way. Can you just go look at the case law on it? And it'll pull up. And it'll like all of a sudden you'll be an expert. in, you know, releases, like security posit releases, whatever the hell it is. Right? You'll be an expert in that one thing because you'll know the case law and you'll write it in such a way where it has to be written exactly and it'll just do that. You know, instead of like what this attorney thought would would would fly.
42:21Unidentified speaker (UNKNOWN): You
42:21David Moses: know, now it's like, okay,
42:24Unidentified speaker (UNKNOWN): the
42:24David Moses: the Michigan Supreme Court, the you know, w whatever it is, whatever county court, they came to a specific conclusion. And here was the actual wording. And so you'll just word it that way right in your clause. Because you
42:38Unidentified speaker (UNKNOWN): know,
42:39David Moses: if it ever comes up and you say, Well, look, the clause says this, and this case is literally directly on point, it says exactly the same thing using exactly the same words that the appellate court used. And the attorneys are like, oh, fuck off. All right, fine.
42:52Jack BeVier: The the thing that I've already seen over reliance on the legal side is like reviewing the thing is that it does a phenomenal job of reviewing the legal document in the four corners of that document. But because it has no real world context until you've given it, like you tend not pe peep people tend not to go to their favorite LLM and start telling it the story of this business relationship and all the things that we've done together and all the emails that have ever transpired between the two parties. And so it freaking it misses that, like, hey, you've got exposure here because you did this and this says that. And like this could be interpreted as a violation of the four corners of the document because it has no idea. Whereas when you're talking to your attorney, a good one at least spends some reasonable amount of time being like, so tell me about your relationship with this person. You got tell give me the story as to everything that's happened. And they can figure out the moments to dive deeper into something that may res you know where where your answer may be like, there's that's what's going gonna to screw you here. Right. They can like a good attorney
43:59can find
44:00Unidentified speaker (UNKNOWN): they can
44:00David Moses: sniff that out.
44:01Jack BeVier: Can sniff that out. and Claude's just sitting there. I mean, maybe they'll build it to the place where they're just like they they ha it interviews and the you interviews you in the same in the same way, and it's still a much cheaper, better version of that. But there's a human intuition aspect of a good attorney that smells where you're gonna get screwed because of their pattern recognition over time. as we call that experience, right? That uh I've already seen people I've already seen I've seen I've like real life examples of people like screwing up through over reliance on on LLMs for you know document review because the you know, the the the LLM gave it fantastic advice, just didn't have any idea as to what that human had done to get themselves in this particular position. And that was the only thing that mattered, you know, like. that was the fatal flaw that Claude had no idea was happening.
44:55Unidentified speaker (UNKNOWN): Or
44:55Jack BeVier: ChatGPT had no idea was happening. Anyway, yeah, can you
44:59Unidentified speaker (UNKNOWN): just say that the other thing is that the other the other
45:00David Moses: side of that is that it will it will oversanitize
45:04Unidentified speaker (UNKNOWN): the
45:05David Moses: strength of your position as well. It will say, oh well yeah but but but this isn't you're not in as strong a position as you think you are because this and this and it's just it just can be flat out wrong because you actually go through and you dig it, dig a bit deeper. And you don't know if it's just kind of trying to agree with you
45:24Unidentified speaker (UNKNOWN): or
45:24David Moses: if it's actually uncovering, you know, what
45:28Unidentified speaker (UNKNOWN): it
45:28David Moses: I I I I have a really good example, but I I I don't I'm not gonna go go into it right now, but I really do think that it's about using it. That's why I'd be fearful if you're an attorney right now who's not using Claude. quit.
45:40Unidentified speaker (UNKNOWN): Yeah,
45:40Jack BeVier: yeah.
45:41David Moses: Sell your practice. There's no point. Right. But an attorney who understands kind of
45:46Unidentified speaker (UNKNOWN): like
45:47David Moses: you're going to go through that conversation with Claude the way you would go through the conversation with your attorney. But Claude can see if you give it the tools to see it, it can see into my Claude can see App Folio. It can see
46:01Unidentified speaker (UNKNOWN): you
46:02David Moses: and you. ubiquitous. It can see everything. Everything in Apfolio, everything in property mailed, everything in every email. It can see it can see my emails. It can see my documents. Right. Obviously that's
46:11Unidentified speaker (UNKNOWN): a
46:12David Moses: gargantuan amount of context. But it can start to ask those questions, right? And you you don't see it yet. Like it's not there yet. But
46:24Unidentified speaker (UNKNOWN): it
46:24David Moses: will get there. It it's less than a year out, I would say. Whatever
46:28Unidentified speaker (UNKNOWN): whatever
46:29David Moses: Fable six is, probably Well
46:31Jack BeVier: they'll never get yeah. As I as I've said as as I said in the previous episode, we'll we'll never get our hands on Fable Six. But we'll see. We'll see today.
46:38David Moses: Because you don't think the you don't think they'll ever release it. They're just gonna they're just gonna if it can cure cancer, why would you bother giving everyone the ability to cure cancer? Why not just cure cancer? Make
46:47Unidentified speaker (UNKNOWN): make all the
46:48David Moses: money yourself.
46:49Jack BeVier: And then and then Fable five 5 will be distilled. and then I'll never pay for anything ever again in a business context. But anyway, the
46:56Unidentified speaker (UNKNOWN): It's very
46:56Unidentified speaker (speaker_2): it's very possible
46:57Unidentified speaker (UNKNOWN): you're
46:57Unidentified speaker (speaker_2): correct. I mean, I'm definitely not going to take the under on that.
47:01Jack BeVier: I do love the idea of the general counsel agent who has access to all the transactions, everything that's happening, all the emails that are being written, all the transcripts that is just constantly monitoring your entire organization and just surfing. surfacing, hey, a little bit of exposure here, a little bit of exposure here. And you could then obvious things getting too uptight, dial it down a little bit, buddy, and tweak your general counsel to the place to the place where you want it to be where it is all seeing and all knowing and reporting to the owner areas of points of exposure such that you can manage out that risk or affirmatively mitigate that risk and not find out about it after the lawsuit hits you and discover. discovery happens and you're like, you said what in an email six months ago? You're kidding me, right? So I do like that use case also. And the fact that that that agent would also have access to all the contracts. So it would have the context of like all of the all of the corporate reps and warranties and affirm you know affirmations that you've made as an entity and make sure that as you operate you're not violating the those things, or when you do violate those things, at least you get a heads up that you did. I really like that idea.
48:08Unidentified speaker (UNKNOWN): Yeah,
48:09David Moses: because you're gonna look every you you create as a business owner, you create exposure every single day.
48:14Unidentified speaker (UNKNOWN): Right.
48:14Jack BeVier: Yeah, of course.
48:15Unidentified speaker (UNKNOWN): Every
48:15David Moses: single
48:15Unidentified speaker (UNKNOWN): day you can't. You cannot do
48:16Jack BeVier: business without getting sued at some point. You can't. Like
48:19David Moses: No. But but and it's just
48:21Unidentified speaker (UNKNOWN): kind of like
48:22Jack BeVier: no, not at scale.
48:23Unidentified speaker (UNKNOWN): It
48:23David Moses: it's not like the criminal that gets caught, right?
48:25Unidentified speaker (UNKNOWN): That's
48:26David Moses: not the first time they do. did that. That's not
48:27Unidentified speaker (UNKNOWN): the
48:27David Moses: second time. That's probably not the fifth time they did that, right?
48:31Unidentified speaker (UNKNOWN): But
48:31Unidentified speaker (speaker_2): they got caught
48:32David Moses: eventually, right? Because they got complacent, because it wasn't big and scary when they did. Think of it that way in that from from not from a nefarious standpoint, but when you make a mistake, when your people make a mistake in business,
48:45Unidentified speaker (UNKNOWN): they're
48:45David Moses: gonna going get to away with it many, many more times than they get caught by it.
48:49Unidentified speaker (UNKNOWN): But
48:50David Moses: then when they get caught, that's where class action lawsuits come from. Right? They come from I made the mistake lots and lots of times, but you add in that nefarious component to it, right? Then it becomes a class action lawsuit. But you take out the you you take out the intent, right? You take out the I've uh, you know, malintent and they're just honest mistakes that you're making that exposing that that are exposing your company. That's that 360 360-degree degree view that I've talked about before. It's just understanding where your exposure is and then and then training around it, right? Because these conversations, this is why your agent should be writing every email. This is why your agent should be sending the text for you. Why? Maybe it's not going to be perfect in protecting you and disclosing everything and doing it. Maybe it's not going gonna to be like that now, but it's gonna be a hell of a lot better than
49:41Unidentified speaker (UNKNOWN): it's
49:41David Moses: like having your attorney on your shoulder before you write that email to say, hey, maybe we shouldn't say this or maybe we should say this, right?
49:48Unidentified speaker (UNKNOWN): I
49:49Jack BeVier: hear you. I'm gonna take the other side of that coin in terms of the communication. I think that the I think that
49:54Unidentified speaker (UNKNOWN): I
49:55Jack BeVier: don't think people talk into a computer like in any context. I think everyone it's always pissed something up it's always pissed us off as humans at some at some level. Now, is there a trigger in my Outbooks send protocol that before the email leaves the Outbook out outbox, it gets reviewed by an agent that is a that that is, you know, the attorney sitting on your shoulder and a little pop up happens and is like, you might not want to say it this way. Think about tweaking it to more something like this. And then the human has to override before the send happens. I yeah, I I'd be I'd be excited about that feature.
50:31David Moses: That's actually that that'd be fairly easy to do.
50:34Unidentified speaker (UNKNOWN): Yeah,
50:34Jack BeVier: it'd be fairly easy to do, I think.
50:35Unidentified speaker (UNKNOWN): Like that
50:36David Moses: little the little paper clip from Microsoft
50:37Unidentified speaker (UNKNOWN): Word. The
50:38Jack BeVier: little stupid paper clip. Yeah.
50:40Unidentified speaker (UNKNOWN): My
50:40David Moses: my
50:41Unidentified speaker (speaker_2): favorite one was like somebody's like, yeah, you know,
50:44Unidentified speaker (UNKNOWN): that
50:44David Moses: paper clip popped up. I see you're writing a ransom note. Would you like some help with this?
50:51Jack BeVier: That's hilarious. Well Dave, hey, I apologize. I got a I got a slammed one today. I gotta cut us off short. We got some good content in here, man. This was I think this is one of the more content packed I mean, you know, you've gotten to a very exciting point that I think a lot of business owners are really interested in in getting to. And you know, you've had it was not a straight line by any stretch of the imagination. So to like
51:13Unidentified speaker (UNKNOWN): have you
51:14Jack BeVier: distilled distill where you landed and how you're dealing with a bunch of issues that you have obviously put a tremendous amount of thought into, I just think is incredibly valuable. So, man, I r really appreciate this conversation. I'm excited to hear how it goes. And uh thanks, man. Much appreciated.
51:28David Moses: We'll see you in a week.
51:31Unidentified speaker (UNKNOWN): Yeah,
51:31Jack BeVier: sounds good, man. So hope everyone enjoyed that episode of Fork.ai. Yeah, Dave, I'll see you next week at Shar in Charlotte at Real M. Investor Roundtable. If anyone's interested in uh, we'll be talking about this and many other real estate investing topics at that uh at that mastermind. So if anyone's interested, thinks they would get value out of that and could add value to the room, don't hesitate to reach out to us. I'm Jack at thedominiongroup.com. And otherwise, I hope everyone has a great week. Have a good one.
52:00Unidentified speaker (speaker_3): Take care.