Published: 2026-08-01

The five-layer "compound agent stack" for a free agent OS

Chapters / key moments (click to jump — plays here on the page)

Julian Goldie describes what he calls the "compound agent stack" — a five-layer way of organising a self-hosted agent setup so that adding a new model does not mean rebuilding your workflow. The layers are a single dashboard with every agent wired in; a shared memory vault (he uses Obsidian, which can run entirely locally); a router that sends each job to whichever model suits it rather than defaulting to the most expensive one; parallel agents worked off a Kanban board by one orchestrating agent; and a loop that feeds completed work back into memory. No terminal work or code is shown — the video's stated method is to describe what you want to a coding agent and have it build the dashboard for you. Much of the runtime promotes the creator's paid package, which ships a pre-built version of the same stack.

Watch on YouTube → · Real-world agent setups · ← Back to News

What you can actually set up from this

Extracted from the video's own transcript — the specifics the original summary left out.

Configuration shown

Layer 1 prompt (dashboard)

Build me a mission control dashboard with Claude and Hermes plugged in. Make it look beautiful.

Layer 2 prompt (memory vault)

Claude, download Obsidian, set up an organized folder with markdown folders and files, and document everything you've done for me.

Layer 4 example task for the kanban board

Find 20 businesses that might need my service and draft a short, friendly message to each on the board before lunch and then review the finished drafts after.

Vault folders suggested

One folder for who you serve, one for what you sell, one for how you talk, and a running log of what you worked on this week. Every agent reads the same vault before answering.

Template shown in the video. Adapt to your project — do not copy verbatim without reviewing each line.

Reproducible steps

Layer 1: foundation
One dashboard with every agent you actually use (Claude, Hermes, OpenClaude, Free Claude Code, a free local model, free APIs). Built by asking a coding agent for it; the speaker says no coding experience is needed. Heavy work runs on subscriptions or free models you already have.
Layer 2: memory (the layer most people skip)
A shared Obsidian vault of plain notes. Obsidian can run fully locally, so paired with local models no data touches the cloud. Effect: weekly emails come out in your voice without pasting past examples.
Layer 3: router
Stop asking which model is best; ask which model wins this job. His split: a free open model for basic tasks, a cheaper model such as GLM 5.2 for middle-tier work, a frontier model such as Fable 5 only when the job needs it. Test the models on your own top three jobs rather than trusting benchmarks.
Layer 4: agents
Parallel workers. He describes Hermes mixture of agents (two agents on the same task and an aggregator that fuses the best of both) and a kanban board where tasks fan out to agent profiles. Use one commander agent that orchestrates the rest instead of prompting each by hand.
Layer 5: loop
Agents do the work, wins are logged to memory, memory improves later answers, and new models are tested and slotted in as new agents or router options rather than rebuilding anything. He says he plugged a newly released DeepSeek in within about 20 minutes.
Where to start
Start with layer 2: install Obsidian (free) and have Claude build a memory vault for your business. Then ask an agent for a simple dashboard and add one agent at a time. Free model options he names include Omni Route and Ling 3.0 Flash.

Gotchas

He says building his own took about 3 to 4 hours a day of continued improvement; the finished system is sold as a zip in his paid community.
The model names, release claims (DeepSeek V4, Qwen 3.8, MiniMax H3 and others) and his leaderboard are the speaker's; none is verified here.
Promotional avatar video with no commands or config on screen beyond the prompts above.