Agent Self-Improvement with Frozen Weights: The Added Value and Limits of Failure Traces
Arkid Mitra (Hopit AI) runs a pre-registered test of whether a frozen agent's harness notes improve more when the note writer sees the agent's failure traces than in a matched search that never sees them.
Ask this paper
Setup. A frozen Qwen3.5-9B agent, with a larger teacher writing its notes. The control runs the same try-and-keep search with the agent's mistakes withheld.
Result. On generated customer-service decisions, failure-based notes beat the matched control by 4.0 points. The search alone accounted for about 60% of the total improvement, so without the control the gain would look 2.6x larger.
No transfer to full episodes. In full customer-service conversations the same notes showed no demonstrated improvement, and a neutral note with no procedure scored higher than learned lessons (exploratory).
Benchmark defects. A blind audit judged 146 of 240 tool-use test items defective. The author recommends a matched no-evidence search, a neutral-note comparison and a blind scorer audit for any self-improvement claim.
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.