AutoResearch
Autonomous ML research harness by Andrej Karpathy. A ~630-line Python repo: `prepare.py` (untouchable, data + utils), `train.py` (edited by the agent, model, optimizer, loop), `program.md` (agent instructions). A coding agent (Claude Code, Cursor, Codex) iterates on `train.py`, each run is exactly 5 minutes, keeps only what beats the current val_bpb. ~100 experiments per sleep cycle.
Strengths
- Minimal: 630 lines of Python total, read the code in 30 minutes, understand the whole pattern
- Validated autonomous loop: agent writes + tests + commits without human intervention, objective metric (val_bpb)
- Works with any coding agent (Claude Code, Cursor, Codex), point the agent at `program.md` and leave
- Karpathy signature + 73k+ stars in days, solid base to study or fork into another domain
Limitations
- Requires an NVIDIA GPU (tested on H100), no full CPU/MPS path, hard hardware entry barrier
- Narrow scope: nanochat small-model pretraining, not a general-purpose harness for any autonomous workflow
- Continuous LLM cost: each loop experiments via a paid agent, the bill climbs fast across 100 experiments
Best for
- Researchers or ML engineers wanting a minimal bench to test autonomous agent iteration
- Studying the "autonomous experiment loop" pattern before transposing it to another domain (prompt tuning, SQL query optimization, etc.)
- Tech leads wanting to concretely demonstrate to teams what an autonomous agent can do overnight