Three scheduled agents, nothing done twice
Long-running agentsSolo founder scheduling headless agents on a home machine
Every night at 1 a.m. a research agent ranks the bug reports and support emails I exported, at 3 a.m. a build agent fixes the top one, and at 6 a.m. a third writes my morning summary. They are three claude -p runs started by Task Scheduler. Before ArcRoom, I passed files between them by hand, and the 3 a.m. agent often missed what the 1 a.m. agent found.
How do I stop scheduled claude -p runs redoing each other's work overnight?
Run each scheduled job as Claude Code headless in the same project folder with ArcRoom on. The 1 a.m. run saves its ranking with complete_task; the 3 a.m. run finds it with find_task instead of redoing it and claims the top item with claim_task; the 6 a.m. run reuses both for the summary. If a run crashes before it saves, the next one only gets the last-run note.
Also asked as: My 3 a.m. agent keeps missing what my 1 a.m. agent found. How do I connect them? · How do I chain several headless Claude Code runs on a schedule? · Can a morning Codex session start from what my overnight agents did?
Each run is Claude Code headless in the same project folder, with ArcRoom on.
- Last-run note
- Claim
- Saved answer
- Done-checks
find_task → reuse: "nightly triage Oct 30" (1:12 a.m.)
Step by step
- Each run is Claude Code headless in the same project folder, with ArcRoom on.
- The research run saves its ranked list with complete_task.
- The build run finds that answer, claims the top item, and must pass the done-checks before it stops.
- The summary run gets the last-run note and the saved answers, and writes what happened.
- In the morning, Codex or Cursor in that folder starts from the same note.
What the agent sees
find_task → reuse: "nightly triage Oct 30" (1:12 a.m.)
1. CSV export drops rows over 10k (6 reports)
2. login loop on Safari (2 reports)
claim_task → "fix: CSV export over 10k"Compared with today's tools
Hand-off files between jobs
Each job writes handoff.md and the next job's prompt says to read it first.
- Setup
- 30 min
- Ongoing
- Fix the chain when a file is missing or stale, ~10 min a week
ArcRoom
Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.
- Setup
- 2 min (npx arcroom init)
- Ongoing
- None
Files are simple and visible; they don't stop a run starting work another run already claimed, and they don't follow work into a teammate's agent.
Where it stops
If a run crashes before it saves, the next one only gets the last-run note. On fresh machines, run arcroom flush at the end of each job so nothing is left queued, as in CI.
Try it in your project
Works with the agents your team already uses, and any agent over MCP or the HTTP API.
More in long-running agents
- Bot-driven sessions that don't redo the last one's workDeveloper who runs claude -p from a chat bot on a home PC
- A three-day migration where no part is done twiceDeveloper leaving Claude Code on a long migration over a weekend
- Coding agents in containers, one per issueTeam running Claude Code headless in containers on its own server