Skip to content
Cameron Mills

All projects

Running unattended AI agents reliably

The Life Tracker agent system

Part of Life Tracker

A manager and a supervisor for headless Claude Code agents, so runs left alone overnight either finish properly or stop in a way I can see and resume.

Code on GitHub

How it works

How one build request becomes supervised agent runs that either finish or stop where I can see them. Select a layer to open it.

Problem

This is the engine of my Life Tracker, the part that runs its AI agents. They are Claude Code agents running unattended on my own Mac, and an unattended agent can hit a usage limit, hang, or stop halfway with nobody watching. The system's rule is that it should take very little of my time to keep running, so anything that needs babysitting does not belong in it.

Approach

Treat every agent run as a managed job, and treat "done" as a claim to check rather than a fact.

Architecture

Every unattended run goes through one agent manager. Each run gets its own folder holding everything the session did, a readable log and the agent's own checkpoint. A watchdog pauses the agent before the usage limit and queues a resume that brings back the whole conversation. A time cap stops runs that overrun, and any abnormal end leaves a crash report with the exact command to resume.

On top sits a build pipeline: one request in, a staged and supervised series of managed runs out. Only the supervisor can mark a stage as passed, and it does so only when the run ended cleanly, its output was written during the run, the output ends with a completion marker, and the agent's last message is not still waiting on something. A stage that fails is re-queued with a note explaining why. Since 14 October a new pipeline also proves it can write code before it starts, and stalls if two build stages in a row land nothing. A queued brief waiting on something that can no longer happen expires after 48 hours rather than being deferred for ever. The scheduler checks in with the supervisor continuously.

What I built

I set the design and the checks each part had to pass, and Claude Code agents wrote the code to that specification. Two of its parts have their own pages: the runner, Steward, and the /build skill that drives the build pipeline.

Results

In its first two weeks almost every run finished cleanly, with a handful killed or timed out. The more useful finding was that "done" from the manager only meant the process had exited and the checkpoint had changed. Several runs that stopped halfway were still recorded as done.

What didn't work

Some runs were recorded done while still waiting on background work, so the next stage started without its input. A written rule in the brief did not stop it. The fix moved the check out of the prompt and into the supervisor, so a "done" stays provisional until the output really exists and the agent is no longer waiting.

Stack

  • Python
  • Claude Code

How these figures are sourced

The outcome figure is agent-reported: it is taken from a test log or report written by the agents that built and ran the system. The status was read from the code repository at commit c909757 on 1 Oct 2026. Every figure and claim on this page is tied to a source file or a CV line in the site's content.