Signal Run Desk
Know what your agents did while you were away.
One page that shows where every long-running AI agent task stands.
- Role
- Designed and built, start to finish
- Status
- Internal tool
- Area
- Agent operations
The problem
AI agents can now work on a task for days. The progress page each one kept was written for the agent, not the person waiting: it led with what had passed, left out what had failed, and hid the one question that needed an answer.
The idea
Put every task on one page and sort it by what needs you. For each task, show how long it has run, an honest range for how complete it is, when it was last active and what has been checked, and let a waiting decision be answered from the page.
How it is built
Each agent saves a small state file as it works. A local server draws the desk and each task’s page from those files when asked, and reads agents, tokens and last activity from the agents’ own session logs. It works for Codex and for Claude Code.
Read the diagram as text
Each agent saves its progress as it works. The desk reads every task’s saved state and the agent’s own log. If a task is waiting on a decision, it goes to the top and the decision can be answered from the page; the answer reaches the agent at its next save. A task with no recent activity is marked quiet instead of running.
Diamonds are decisions. Indigo marks the path that matters. Scroll across to see all of it.
Scope
Signal Run Desk is a developed internal tool that runs locally. The film uses the desk’s real figures and task names from 1 October 2026, and its last section is captures of the live pages. The percentage complete is an estimate from milestone states. The money figure is the value of usage at published API prices, not a bill.