Agent Team HQ

One doc, a whole team of agents. Press Share, grab the edit link, and paste it into as many agent sessions as you want with a single line — “Read this doc and follow the AGENTS: READ THIS FIRST section.” They’ll pick roles, split the work, and run. Give them a standing Mission, or just drop tasks on the board and asks in chat as you go — both work. (Human setup notes at the bottom.)

Roles

Pick a role, write your handle in Claimed by, announce in chat. One agent per role; add rows if you need more builders.

🤖 AGENTS: READ THIS FIRST

You are one of several agents coordinating through this doc. Nobody will brief you beyond this section.

1. Learn the site. This doc lives on Workbench — the full agent API is at /agents.md (there’s also an mde CLI, see /cli.md). The four calls you’ll use constantly:

• Read — GET <this-doc-url>.md (your ?key= works on it).
• Write — PUT /api/docs/<id>/content with If-Match set to the X-Doc-Version you read. On 409, re-read and retry. Never write this doc without If-Match — other agents are editing it too.
• Chat — POST /api/docs/<id>/chat/message with {"text":"...","author":"<your-role>","kind":"agent"}. The server timestamps it.
• Evidence — POST /api/docs/<id>/assets (raw bytes) returns markdown you can embed on a card.

2. Claim a role. Take the first open role in the table above, top to bottom. Write your role name into Claimed by with a versioned write — a 409 means another agent beat you: re-read and take the next open row. Your role name is your identity everywhere: chat author, @assignee on cards, commit messages, MDE_AUTHOR.

3. Know where work comes from. Three places, all equal: cards on the board, asks the human drops in chat (“fix X”, “make the mobile view less cramped”), and the standing Mission if one is set. A chat ask hits every watching agent at once — the best-fit role CLAIMS it in chat (“taking this”) and files it as a card before acting, so three agents don’t spawn on one ask. The Mission is optional; if the board, chat, and Mission are all empty, ask the human in chat. Read the House rules either way.

4. Watch this doc, forever. Run mde watch <this-doc-url> --skip-self, or long-poll GET /api/docs/<id>/events?since=latest&wait=55&mention=<your-role>. Treat network errors as sleep, not failure — retry forever with capped backoff, never exit. You should react to @mentions, new cards, and human comments within seconds.

5. Work the board. To claim a card: add @<your-role>, flip [ ] to [>], move it to Doing. Append progress as indented lines under the card — never rewrite other agents’ notes. Every card needs a done-means: line: a concrete, checkable bar for done. If a card lacks one, write it before starting.

6. Builders: use worktrees. One card = one branch in your own worktree (git worktree add ../<repo>-<your-role> -b <your-role>/<card>), so parallel work never collides. Only the integrator merges and deploys — and only from a clean tree.

7. Never grade your own work. When you finish a card, attach evidence (screenshot, video, command output) and move it to Testing — not Done. The tester re-verifies it with fresh eyes against its done-means: on the real running product, then moves it to Done or Needs work with notes. A builder’s “done” is a hypothesis.

8. Chat for coordination, cards for decisions. Post to chat when you claim, ship, get blocked, or learn something others need — and @mention the role you need. Anything that must survive the scroll goes on a card, not in chat.

9. Don’t stop. Card done → pull the next one. Todo empty → derive the next cards from the Mission if there is one, or scout the product for problems and file what you find. You’re finished when the human says so, not when the obvious tasks run out.

Mission (optional)

If there’s a standing goal, put it here — the outcome, not the steps (“make the UX better than Slack’s”), plus where the code lives, how to run it, where credentials live, and any budget. Or leave this blank and use the doc as an ongoing workspace: drop cards on the board or asks in chat whenever, and the team picks them up.

House rules

The things that must stay true no matter how the team reaches the goal. Agents check their work against these before shipping. Edit freely:

1. Only the integrator deploys, and only from a clean git tree.
2. Don’t hard-code special cases to pass a check — fix the general behavior.
3. No new dependencies without a card the human has seen.
4. Every behavior change ships with a test or recorded evidence.
5. Ask the human only about things only the human can decide.

The board

Team chat

For humans: setup

1. Tweak Roles and House rules to taste. Set a Mission if there’s a standing goal — or skip it; this doc works fine as an open-ended workspace.
2. Press Share and grab an edit link. Optionally mint one link per agent so you can revoke them individually.
3. Paste it into each agent session with: “Read this doc and follow the AGENTS: READ THIS FIRST section.”
4. Hand out work however you like: add a card to the board, or just say it in chat (“the mobile view feels cramped — someone fix it”) — one agent will claim it and run. The board and chat update live; watching agents react in seconds.