The idea in one paragraph.
Drop the Godel machine's demand for proof. Let agents rewrite their own code, validate each change empirically on benchmarks, and keep an archive of all variants as stepping stones.
THE MACHINERY
Self-modification
The agent edits its own repository to produce a new agent.
Archive of agents
Keep every variant, not just the best, as stepping stones for later branches.
Empirical validation
Benchmark score replaces the formal proof of benefit.
Open-ended exploration
Sample parents from across the archive to explore many paths in parallel.
What the paper reports.
Findings are reproduced as concise corpus summaries, with their original confidence labels. They are not independent replications.
Improves itself on SWE-bench from 20.0% to 50.0%, and on Polyglot from 14.2% to 30.7%.
Inspect a source excerpt
increasing performance on SWE-bench from 20.0% to 50.0%, and on Polyglot from 14.2% to 30.7%Read the passage in context on arXiv ↗
Discovers its own harness improvements - better edit tools, long-context management, peer review.
Significantly outperforms baselines without self-improvement or open-ended exploration; all experiments were run with sandboxing and human oversight.
LIMITATIONS / CORPUS READING
Where to keep your judgement.
- Assumes benchmark score tracks real coding capability - the paper names this as a key assumption.
- Rewriting training scripts or training a new foundation model is explicitly left as future work.
- Objective-hacking risk is real enough that the authors ran everything sandboxed with human oversight.
- Anticipates precisely the failure Prime Agent later documents: 'if evaluation benchmarks do not fully capture all desired agent properties (e.g., safety and robustness), the self-improvement loop could amplify misalignment over successive generations.'
Why it belongs in this atlas.
Rung nine: the agent rewrites the harness. The talk's word for the loop is meta-harness - a harness whose output is harnesses. Note that safety stops being optional here.