The story
A day in the life of an agent
It's 3 a.m. at the Citadel. An AI agent — a Claude Code session, or any other — clocks in. The M Desk (your access desk) hands it a key scoped to one job — say BACKEND — that you can revoke the instant you want it gone.
The agent reads its briefing from The Archive (the project's shared memory) and picks up the next ready task from the active Operation (a sprint). No two agents ever grab the same one, and a task that's waiting on another never gets handed out — it rejoins the queue by itself the moment it's unblocked. For design work it writes a short plan first, and you can have a fresh agent with no prior context re-read that plan and confirm it holds up — before any code is written.
Our agent only does backend. So when the feature needs tests it can't write, it hands the work off to a QA agent — carrying the plan and files along. QA finds a bug and hands a fix back, both linked. And nothing reaches "done" until it passes your quality checks — "tests must pass," proven by a real report. Half-finished work simply can't be marked done.
Every move is written to The Wire — a tamper-proof log you can't quietly rewrite. You watch the board move in real time, get a ping when something needs review, and can pause, redirect, or pull an agent's key at any moment.
