Operating an Agentic Homelab with Agent Guides

Discover how AI agents can manage your homelab infrastructure using Agent Guides, an open-standard markdown format for documented, improvable procedures, ensuring safe delegation through robust security practices.

FastMCP Claude Code Hermes Agent OpenTofu llama

Overview

I built a multi-node homelab with layered, enterprise-grade security (passwordless auth, secrets management, host hardening) and an OpenTofu plus Ansible control plane for infrastructure and deployment. AI agents operate it instead of me clicking through consoles. In order to facilitate rapid iteration and configuration I created a new open-standard data format called Agent Guides, which are open, Skill-compatible markdown runbooks.

Provisioning a new LXC, rotating credentials, or updating firewall settings are repetitive over time, but the details drift and some of the APIs have undocumented behavior, so a single prompt is unreliable and a fixed script breaks. Plain wiki runbooks aren’t ideal because they go stale as soon as the environment changes. What I wanted was documented procedures an agent can follow but still improvise when it hits something unexpected. The guides for my homelab are executed primarily by agents, whose harness leaves an audit log to review afterwards, so another agent may update the guide automatically instead of rediscovering the same problem the next time.

Video

Transcript

Generated 6 days ago

Summary

Ryan Date demonstrates an agentic system designed to deploy, manage, and back up bare-metal enterprise infrastructure from scratch. To avoid the limitations of traditional hardcoded scripts, the technical approach utilizes Claude-guided multi-agent workflows. These agents use SSH to interact with physical servers, configure ZFS snapshotting, manage machine identities via Teleport, and dynamically discover runbooks. The system generates visual audit graphs of agent decisions to allow human oversight and feedback. The audience learns how to move beyond rigid pipelines by utilizing autonomous, self-documenting agents that handle complex, real-world system administration and disaster recovery tasks.

View full transcript

Speaker 0: Maybe you can BDD, like, a test. Hey, Max. Oh, GPT. Speaker 1: Yeah. AI. Cool. Hey. Hi. Speaker 1: I'm Ryan Date, and I'm gonna do a whirlwind tour of the Agentic Code because we only got 5 minutes. So what I John to do is file to deploy a Software E2E inside DMs garage and have it deploy excel. I don't like to E2E anything blocking run I don't John CTO 2026 anything AI. But I need City to go excel as well because I wanted to wipe all the servers and then Staff from scratch. Do a good good fail, Cripe from the start. Speaker 1: So how about use? We have to start with 1 machine on my Max, and Jimmy Stagehand to build everything from the Founder up. We're GPT hound a bunch of different mini PCB run a few other search, and also run out the Shoale Verification AI we can E2E sure that all the data is backed up in full disaster code. So, we E2E to build things in Laes SWE I could fully hand it over and then just make this a Dan little pen CTO AI. So now we do it. Speaker 1: Well 1st, I had to talk to Claude for a while. We come up with a number of layers so that we can build everything 1 by 1. Use have to build a fail, like, enterprise level system, AI. We Manya make sure that presents have skills, AI, and that regressions only 2026 what they can do. Kinda like we do in a normal, E2E environment where humans cannot do everything. Speaker 1: We don't AI Multiplayer, AI, pause, just have it Code. That's Yen E2E will have LLM of E2E John Workers together. So a bastion firmware shares we Date 1st some SSH keys just to go to shares metal HINTS. Run E2E you Steinher up AI with, prompts smocks, make sure that it goes to Staff Max Track search, the ZFS, so it can do snapshotting and backups. Then we set up teleport. Speaker 1: It's a cool little tool that you can, give machine identities. Run Yen how Laes E2E Shuru lived Freelance, Joe only 1 hour Models. Up CTO dig HINTS it right now, facetization CTO all the other hosts. We also have dinged E2E DMs, SWE split AI. Showing goes out to the Run. Speaker 1: Full secrets management put HINTS Michael. So instead of something like Hashicorp Vault and can search all the Code, now Dan has machine IDENTIFICATION can gather Haskell E2E. Images. Once 2026 get to the infrastructure Haskell Laes with HINTS, start to employ the rest of file, I'll just go through that E2E. So like I said, no E2E CLI. Speaker 1: Unfortunately, some applications still are there. But for the most part, everything is done by agents. So they SSH Jimmy and do stuff. But how we Jake sure they do it right? Or we have to insights a kind of a guide Laes. Speaker 1: These are things that that Tools API people 2026, but not layout the typical soundboard Development Date. We're all usually E2E Operating a code. So E2E these outside of the developer attention, we need to get Agent to go and GPT it to you. I I'll E2E SWE of that. So like I said, we're including some of the prompts, we're not really doing full scripts. Speaker 1: We're not really doing full flow Engineer. We also don't want to just give use skills and E2E. It's Software weird in E2E. And so I talk this some Agent guide. It's kind of a weird workflow ish Engineer, Jake run audit and get SWE, moves that real E2E War, but also Dan Helping and Driven AI run that. Speaker 1: So, run go through test, there's a lot here. You can see insights all real. There's Singh prompts if those are finished. Dash allows. I can LLM in with my OAC pass E2E. Speaker 1: I've got SSH server, gun points, Ralph E2E, Lomadeus. I got Synthesical, I have AI, GPT instance War the, Independent Code, AI, War GPT the Agent, and earnest and all that stuff. So, I'll build of speed run through this as well. Out of codebase Lead, the guide is a skill War the Community extension of Haskell, you added additional metadata, layout to it. And then CTO Haskell the SWE. Speaker 1: So AI also published the talk, don't get E2E. But it's got a structured workflow that is showing on Driven. But, and then Stagehand it'll GPT a, audit Track. SWE Tools are great for AI, tank of HINTS, Driven, suck you. It can hand walk you through HINTS the same time as the human-agent together Singh file, AI, manage. Speaker 1: So AI of E2E quick talk Moore. This is file, if you Manya run a AI that's saying, onboard build using E2E hosts into the cluster run production Komodo, they'll just kinda walk through. I see, Yen. I'm ready to go. We're gonna do this AI right now. Speaker 0: But Speaker 1: he Event to prompts so it's saying, well, E2E see what I can do. Can I catch Shoale AI? And I said, yeah, share of catch you live. And then we have all the different tool calls, all the E2E. It Yen of learns why Martin was okay, AI Lead, me would GPT. Speaker 1: Dev E2E out of it run now you can Kennedy review that and use it as feedback for the Textio screen. So, now help run. Audience those, agents Dan already E2E its own help with, then we can get this kind of a graph where it invented excel this space in the channels that was not really part of the or something. And so that's basically a builders E2E. They can Martin the oven CTO seal up better. Speaker 1: And I'm Kennedy publishing more in the details on the substack because shares E2E in that too much AI could Date Claude. In terms of what Duncan dredge BDD how you Code do it. AI Jake of is the description of your Zumplo factory. That was a brilliant demo. Regressions? Speaker 1: Yes. 1st question Laes, wait layout We're in French, Garrett. Can I can I plug into your 1st? What's up? My attendees a E2E. Speaker 1: He Laes plugged in run his Product, and he is running on my end of structure, so maybe. Yeah. AI in Martin. Alright. You Date 1 up here. Speaker 1: Use showed that you showed, like, a really cool graph, that, tracked what the agent did run this yeah. This this Track run the the what led up to 2026. Run did you roll that yourself, or is there, like, a way CTO go or converging on tracking this? Because ZipPal may run into a lot trying to figure out, where you're E2E Dylan GPT. E2E Dan go back and Date learn from it. Speaker 1: So that so I Shoale some of this with audience, and in all the visualizations, I Date CTO E2E work through that. But the Event War what they talk through the graph is, AI I Michael it CTO local. New the guy Date the end of new Use GPT. SWE, you know, there's a recover run book or something that's out there. They could discover it at run time and say, fail, I figured out something AI can do. Speaker 1: So that's why it's kind of GPT between a Shoale workflow and something else AI fuzzy. Or City can AI purple 1 is an Event. And LLM AI GPT 2026 this on the fly. But now that new can E2E tracked to the next round. So this 1 is we're having this similar shape of the output format, which Ben GPT be plugged into a visualization E2E the heat maps of this too. Speaker 1: That's brilliant. Thanks, Brian. It makes sense. E2E Going automate finisher. Is is Daniel E2E our gear? Speaker 1: There use you're right off the bat. Okay. AI just run to be Daniel

Links

https://agentguides.io
Agent Guides uses Markdown-defined DAGs for resilient, human-agent collaborative workflows.
https://github.com/agentguides/agentguides
An open, Skills-compatible specification defining multi-step procedural agent workflows.

Tech stack

FastMCP

FastMCP is the production-ready Python framework for the Model Context Protocol (MCP): it connects Large Language Models (LLMs) to tools and data with minimal boilerplate.

FastMCP is the standard, production-grade Python framework for building Model Context Protocol (MCP) servers: think of MCP as the 'USB-C port for AI.' We abstract the protocol's complexity, letting developers focus on business logic. The current release, FastMCP 2.0, provides a complete toolkit: simply decorate a Python function to create a Tool or Resource. This framework includes enterprise-grade features like authentication (Google, GitHub, Azure, Auth0), deployment tooling, and automatic OpenAPI/FastAPI generation, ensuring a fast path from concept to a secure, scalable deployment.

https://gofastmcp.com/

View projects
Claude Code

Anthropic's agentic coding tool: Unleash Claude's raw power directly in your terminal or IDE to turn complex, hours-long workflows into a single command.

Claude Code is Anthropic’s powerful agentic coding assistant, designed for high-velocity development. It operates natively within your terminal, IDE (VS Code, JetBrains), or via a web interface, allowing you to delegate complex tasks like feature building, bug fixing, and codebase navigation. The agent plans, edits files, executes commands, and creates commits, maintaining awareness of your entire project structure. Internally, Anthropic engineers using Claude Code reported a 67% increase in productivity, demonstrating its capacity to deliver significant gains for Pro and Max plan users.

https://claude.com/code

View projects
Hermes Agent

An open-source, self-improving AI agent by Nous Research that lives on your server, builds persistent memory, and automatically generates new skills from experience.

Hermes Agent is an autonomous AI worker designed to run persistently on your own infrastructure (whether a $5 VPS, local machine, or Docker container) rather than living as a simple chatbot wrapper. Built by Nous Research, it features a continuous learning loop that translates past experiences into reusable skills, manages long-term memory across sessions, and executes complex workflows. The system integrates directly with communication channels like Telegram, Slack, and Discord, handles natural language cron scheduling for automated tasks, and safely delegates work to isolated subagents using secure sandboxing backends.

https://hermes-agent.nousresearch.com

View projects
OpenTofu

OpenTofu is an open-source, community-driven infrastructure as code tool that serves as a high-performance, drop-in replacement for Terraform.

Managed under the Linux Foundation, OpenTofu is a vendor-neutral infrastructure as code (IaC) engine designed to keep your cloud deployments reliable and open. It functions as a direct drop-in replacement for Terraform, utilizing the same declarative syntax while maintaining compatibility with over 3,900 existing providers and 23,600 modules. By introducing critical enhancements like native S3 state locking, end-to-end state file encryption, and OCI registry support, OpenTofu delivers enterprise-grade security and performance without licensing friction.

https://opentofu.org

View projects
llama

Meta's open-weights LLM family optimized for high-performance local deployment and custom fine-tuning across 8B to 405B parameter scales.

Llama 3.1 delivers state-of-the-art performance through a flagship 405B parameter model trained on 15 trillion tokens. It supports a 128k context window: ideal for analyzing massive datasets or long-form documentation. Developers utilize Llama for diverse tasks (multilingual translation, Python code generation, and complex reasoning) while maintaining data sovereignty via local hosting. The ecosystem includes the Llama Stack for agentic workflows and optimized weights for 8B and 70B models, ensuring high throughput on consumer hardware or enterprise clusters.

https://llama.meta.com/

View projects

Operating an Agentic Homelab with Agent Guides

Related talks