Four agent lifecycles, one markdown file format
Autonomous agent management for Claude Code, where every agent is a markdown file you can read, grep and edit.
A background daemon runs agents defined as markdown with YAML frontmatter: configuration in the frontmatter, instructions in the body. Four different execution models share that one schema. Scheduled agents run on cron. Watcher agents run a cheap check script and only wake Claude when the condition is met, which is the design that keeps the cost down. Persistent agents hold long-running sessions with heartbeats and checkpoints. Issue agents take a Linear or GitHub link over Telegram, spin up Claude in a git worktree, open a PR, and remediate reviewer feedback until it is clean.
That last one has no natural stopping point, so the termination condition has to be explicit rather than emergent.
Run logs, memory and costs are written back as markdown with backlinks. A database would query faster and make the whole system opaque. Being able to open a run log in an editor, or grep six months of them, is what makes an autonomous system supervisable rather than merely trusted.