A 𝟰-𝗺𝗼𝗱𝗲𝗹 𝗿𝗲𝘃𝗶𝗲𝘄 𝗼𝗳 𝗺𝘆 𝗛𝗲𝗿𝗺𝗲𝘀 𝗪𝗼𝗿𝗹𝗱 𝗘𝗻𝗴𝗶𝗻𝗲 design 𝗽𝗹𝗮𝗻

I'm building the Hermes World Engine: a world that stays the same, with characters brought into it, so I can make animations that teach Hermes. This week I wrote the plan to expand it into a studio, and I wanted it reviewed by more than 1 model. So I set up Mixture of Agents in Hermes Agent: Kimi K3, DeepSeek V3.2 and GPT-5.6 Terra advising, Claude Opus 5.5 making the call. I gave the same plan to GPT-6.1 Sol alone. Sol rated my routing sound. The team found the missing boundary: the image route assumed approved assets already existed, and nothing said when a route was complete.

The first 2 runs failed before I found the fix, and the fix was partly an ICM fix. The plan, the review rules and the video itself all live in ICM folders, across 2 workspaces.

━━━━━━━━━━━━━━━━━━━━

𝗧𝗛𝗘 𝗪𝗢𝗥𝗟𝗗 𝗘𝗡𝗚𝗜𝗡𝗘 𝗜𝗖𝗠

• The workspace sits on my Hermes server. The root carries AGENTS.md, CLAUDE.md and CONTEXT.md as the router, plus IDEA.md and README.md.

• The stages: 00_setup, 01_spec, 02_render, 03_file, 04_animate. Today's route runs from a video brief to brief approval, a motion spec, a render, a consistency review, then acceptance and filing.

• world/ is the canon: the characters (Hermes), and the environment with 10 locations, from Aether Falls and Crystal Citadel to the Gateway Pavilion, plus its core, forms, materials and light.

• _reference/ holds the rules every stage reads: the world bible's template, the drift checklist, motion rules, spec rules, the setup interview, the routing map, and the studio extension design that got reviewed.

• runs/ and tests/ hold what each render produced and the checks on it.

The review rules went into _reference/ as well: judge outcomes, owners, handoffs and acceptance, and leave the tools alone. That one file is what got the advisers reviewing my plan instead of going looking for my tools.

━━━━━━━━━━━━━━━━━━━━

𝗧𝗛𝗘 𝗣𝗥𝗢𝗗𝗨𝗖𝗧𝗜𝗢𝗡 𝗟𝗜𝗡𝗘 𝗜𝗖𝗠

• 9 stage folders, 01_gameplan to 09_blog. Each holds a CONTEXT.md contract: what it reads, what it must not load, what it makes, and where I check it.

• A stage only runs when the stage before it says approved. I approve in chat. Nothing approves itself.

• 1 config file holds my answers once: brand, colors, voice, links. When 2 files disagree, the config wins.

• _reference/, checks/ and skills/ hold the rules, the scripts that check the work, and how each step gets done.

• Each video is 1 folder. This one ran: a live recorded session, then the gameplan, a cleaned script from my own transcript, the spec, the voice in ElevenLabs from that script in 18 stitched pieces, the storyboard, a 90 second approval render, 3 cuts rendered in code, the package and the blog post.

━━━━━━━━━━━━━━━━━━━━

𝗪𝗛𝗔𝗧 𝗜 𝗧𝗢𝗢𝗞 𝗙𝗥𝗢𝗠 𝗜𝗧

Both workspaces taught the same thing. A plan without boundaries drifts, and so does a review. Give each one a folder, a contract and a box.

Thank you to

@Jake Van Clief

and

@David McDermott

for ICM, and everyone in Clief Notes.

The full write-up: https://agentsasfolders.ai/blog/mixture-of-agents-hermes/

More information at https://agentsasfolders.ai

5 comments

A 𝟰-𝗺𝗼𝗱𝗲𝗹 𝗿𝗲𝘃𝗶𝗲𝘄 𝗼𝗳 𝗺𝘆 𝗛𝗲𝗿𝗺𝗲𝘀 𝗪𝗼𝗿𝗹𝗱 𝗘𝗻𝗴𝗶𝗻𝗲 design 𝗽𝗹𝗮𝗻