ABREKA: an autonomous research agent that runs for days, one experiment at a time
generativist · x · 2026-11-15
generativist (jbn) shares his open-source ABREKA—a CLI tool automating the research experiment loop of propose → implement → test → run → evaluate, running autonomously for hours or days while building on its own results. Design points:
- Experiments as the unit of work: every action yields an experiment with hypothesis, method, metrics, and findings;
- Fully automated by default, human-in-the-loop optional: abreka run 24h vs --step single iteration;
- TDD enforced: code must pass tests before experiments run;
- Full auditability: every agent session transcript and structured metric is saved.
He also notes it's been remarkable to watch how each new model adheres to and resists the task tree differently, and which ideas get followed. The internal team version is now coupled to private infrastructure; the public one isn't updated.
More from coding & agent
- An AI agents developer roadmap: programming, math, and LLM fundamentals — blaizedsouza · 2026-11-15
- Personal AI agents need their own social network to coordinate for you — StewartalsopIII · 2026-11-15
- Google's PaperOrchestra uses 5 AI agents to write a full paper in ~40 minutes — alex_verem · 2026-11-15
- The four-layer playbook for agents that repeat the same mistake all night — blaizedsouza · 2026-11-15
- How I rebuilt my daily schedule around AI agents: overnight runs and high-intensity check-ins — vinvan · 2026-11-15
- Haskell dev vibe-codes himself a better ghcid, and so can you — tomjaguarpaw · 2026-11-15