I built MartinLoop because AI coding agents are becoming useful enough to touch real repos, but most agent runs still lack basic operational controls.
The core question I’m trying to answer is:
“Can another engineer audit this run later?”
MartinLoop is open source and currently focuses on:
- hard budget caps
- JSONL run records
- audit trails
- failure classification
- test-verified completion
The harder problem is not just stopping an agent when it hits a limit. It’s stopping cleanly at a safe halt boundary with an actionable diagnostic, so the repo is not left in a half-finished state.
I’d love feedback from people using Claude Code, Codex, Cursor, Devin-style agents, or custom coding loops.
New improvement coming with oss core 2 and 3 releases in coming weeks! Headless features for agent2agent interactions, self learning with Sansa, opentelemetry, smart model routing, its everything you would need in a complete solution and its built already. Phased rollout is being scheduled. Try the free open sourced core, new benchmarks coming!