The AI agent that runs on GitHub Actions and wants you to ignore it

What it does

Aeon is a TypeScript framework that parks autonomous AI agents inside GitHub Actions. You pick from 156 skills—research digests, PR reviews, market alerts, social posts—and set a schedule. The framework dispatches workflows, scores its own output, detects when a skill degrades, and patches failing skills without intervention. It is explicitly built for background work you want finished while you are asleep, not for live pair-programming.

The interesting bit

By treating GitHub Actions as both runtime and infrastructure, Aeon eliminates servers to maintain and processes to keep alive. It can also spawn child instances of itself through fork-fleet and spawn-instance, spinning up isolated specialists—say, one for crypto monitoring and another for repository hygiene—that operate from their own forks without sharing secrets.

Key highlights

• 156 independently schedulable skills across research, development, crypto, social, and meta-categories.
• A self-healing loop that monitors skill health, scores output quality, and fixes broken skills automatically.
• Persistent memory across runs and reactive triggers that fire when conditions change, not merely on a cron cadence.
• An optional “soul” system: drop SOUL.md and STYLE.md files into a soul/ directory to make every skill write in a consistent voice.
• A local Next.js dashboard that delegates all heavy lifting to the GitHub CLI; the actual agent logic runs entirely inside Actions using a Claude subscription or Anthropic API key.

Caveats

• The dashboard is tightly coupled to the GitHub CLI and cannot manage secrets or dispatch workflows without it present on the same machine.
• Spawned fleet instances do not inherit secrets, so each fork requires its own manual key configuration before it can run.
• The authors concede the tool is not for interactive coding—you still want Claude Code for that—meaning Aeon is strictly a batch automation layer.

Verdict

Worth a look if you have recurring tasks—morning briefs, repo hygiene, market digests—that do not need you in the loop and can run on infrastructure you already pay for. Skip it if you are looking for an interactive coding companion; the README itself admits Claude Code still wins there.