# ByteHive Supervisor

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Dat Dang](https://cerebralvalley.ai/u/tdat)
- **GitHub:** https://github.com/tdat-dev/bytehive-supervisor
- **Demo video:** https://youtu.be/9Ra9-X84jVM
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2391

ByteHive Supervisor catches AI coding agents that say they are done but aren't, loop on one step, or are about to run something dangerous. It gives you one light, one plain sentence and one button per agent.

I run several AI coding agents at once, and reading their terminals does not scale. An agent says "All done" with a TODO left and a failing test. Another repeats one step over and over. Another asks to force push, and the prompt looks like every other prompt.

Supervisor is a separate process that watches every agent running in ByteHive, my desktop app for agent CLIs.

- Fake-done verdict: checks a "done" against the git diff, a TODO and stub scan, the test results on the agent's screen and its board card. Red with the reason, then Send back and Recheck. An agent with no test run can still go green; the panel lists "No test run seen".
- Stuck radar: red when an agent repeats the same screen or goes silent, with one rescue button.
- Same-file alarm: two agents changing one file, across worktrees of a repo, with a button that tells the second to stop.
- Plain questions and a risk gate: reads what an agent is asking, marks the safe answer, and holds force pushes, recursive deletes and hard resets in red until a person answers.
- Every button is one line in the workspace outbox, which ByteHive types into exactly that agent. Nothing is sent unless a person presses a button.

All code in the repository was written 26 Oct to 4 Nov 2026, and 174 tests pass. ByteHive, the app it plugs into, existed before the hackathon and is not part of the entry. In the demo, six of the seven agents are scripted so the run can be repeated; one is Muse Code really doing its task. The repos, tests, messages and Muse's sentences are real.

Not done yet: the Restore button is not built, and dollars per agent is tested but not wired to real usage.

Website: https://tdat-dev.github.io/bytehive-supervisor/

## More from Meta Global AI Developer Hackathon

- [StatWise Innovators](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2388)
- [ReproTrace AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2389)
- [SmearLens](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2390)
- [void catalyst](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2392)
- [Arosa](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2393)
- [Croissants enjoyer](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2394)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2391. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt