rsi-loop

Recursive self-improvement for research agents evaluates and selects agent rewrites against a fixed task battery. It proposes mutations, tests candidates under a fixed token budget, and scores them on held-out private data to ensure strict improvement.

Benchmark agent rewrites by testing candidates across public task families under a fixed token budget, then scoring strictly on held-out private data. Reject any candidate that fails verifier audits or does not strictly beat the incumbent agent.

Adversarial verification prevents reward hacking by applying verifier audits to candidates after private scoring. Candidates failing these checks are rejected, ensuring selected agent rewrites represent genuine improvements rather than exploited metrics.

The benchmarking task battery includes bin-packing, tabular-classification, and instruction-routing tasks. These multi-step tasks evaluate candidate agent rewrites and obtain comparable scores under a fixed token budget during the improvement process.

Use strict separation between public and private data when running recursive self-improvement loops. This prevents overfitting by evaluating candidates on public task families first, then scoring strictly on held-out private data to verify genuine performance gains.

To perform an rsi-loop step, generate a focused mutation of the incumbent inner agent based on the run ledger, evaluate it on public tasks under a fixed budget, apply private scoring, run verifier audits, and select only candidates that strictly beat the incumbent.