docs(orchestrator): document that agents cannot complete cards

Add an "Implementation Board" section to the root AGENTS.md (auto-loaded as
context by every pi run) stating the hard rule: an agent may never mark a kanban
card done — only a human can. Agents leave finished work in-progress with a
summary comment for review. The code-level enforcement lives in the API.

AGENTS.md

@@ -48,6 +48,20 @@ cargo clippy            # Lint
 - **V1 is single-player** with local persistence. Future multiplayer servers will be added as an optional enhancement.
 - **Naming follows RFC 344**: see `apps/game/AGENTS.md` for Rust/Bevy conventions.
 
+## Implementation Board (Agent Orchestrator)
+
+The docs site's Implementation Board (`apps/api` + the `KanbanBoardPage`) can
+launch autonomous `pi` agent runs against individual cards from isolated git
+worktrees, streaming progress back to the UI.
+
+**Hard rule: an agent may NEVER mark a kanban card `done` (completed) — only a
+human can.** When an agent finishes its work it leaves the card in `in-progress`
+and posts a summary comment (plus links to any docs it updated); the operator
+reviews the diff and moves the card to `done`. Agents may move cards freely
+between `backlog`, `todo`, and `in-progress`. This is enforced in code: the
+agent-facing card-move endpoint (`PATCH /api/internal/cards/:id`) rejects `done`
+with `403`.
+
 ## Conventions (TS apps)
 
 - TS apps use Tailwind CSS v4 with custom design tokens from `packages/ui`