Agent Self-Improvement with Frozen Weights: The Added Value and Limits of Failure Traces
Organizations: Hopit AI · Overlayy
Abstract
We evaluate whether an AI agent's harness, the notes around its model, can learn from the agent's failure traces beyond a matched search that never sees them. It can, but less than it looks. We tested an open 9-billion-parameter model with fixed weights; a larger teacher model wrote its notes. On generated customer-service decisions, where both note writers were told the target rules, notes based on the agent's mistakes improved accuracy by 4.0 percentage points over notes from the same process with the mistakes withheld. That process alone, trying and keeping better-scoring notes, reached about 60% of the total improvement. In full customer-service conversations the same notes showed no demonstrated improvement. Separately, a strict screen of single lessons admitted two customer-service lessons, still awaiting tests on new data, and one tool-use lesson, written without seeing the agent's mistakes, which showed no demonstrated gain on new tool-use tasks. A neutral note with no procedure scored higher than the learned lessons (exploratory). A blind audit judged 146 of 240 tool-use test items defective.
Figures & tables
| Test | Estimate | Reading |
|---|---|---|
| decisions: trace lessons no-evidence search, accuracy (registered primary) | +4.0 [+2.3, +5.7] | passes superiority; point estimate below the 5-point worthwhile gain |
| Official retail: trace lessons standard harness, pass^1 | 1.3 [ 5.9, +3.3] | no demonstrated improvement |
| gate: “address change first” / “list only the first change” / both together | +4.7 [+4.0, +5.4] / +2.7 [+2.0, +3.4] / +7.4 [+6.3, +8.4] (CrI) | passes the screen; no held-out test was run |
| BFCL gate: best lesson (from the no-evidence search); 0 of 8 trace lessons pass | +4.9 [+1.9, +8.0]; audited +3.9 [ 0.4, +8.3] (both CrI) | passes the screen |
| BFCL held-out: standard harness that lesson (positive favours the standard harness) | +0.31 [ 3.12, +3.65], Holm 1.000 | null; did not replicate |
| BFCL held-out: neutral note standard harness (descriptive) | +4.5 [+0.7, +8.4] | registered descriptive secondary; higher observed score, unadjusted for multiple tests; no confirmatory claim |
LLM Agent Self-Improvement
Large language model agent self-improvement studies how agents use experience and feedback to improve their behavior, tools, and workflows.
See all 296 papers