Reason Wide, Not Deep: Amortizing the Reasoning Premium into Distilled Skills

Introduces passive skill distillation, a lightweight way to amortize test-time reasoning across repeated agent tasks. A coding agent analyzes 35–50 trajectories from a domain, extracts recurring failure patterns and successful procedures, and compiles them into a compact 40–130-line natural-language skill. The skill is then placed in a non-reasoning model's system prompt: no model weights are updated and no fresh environment rollouts or per-instance search are required at deployment.

Across ALFWorld, SpreadsheetBench-Verified, and the telecom and retail domains of τ²-bench, distilled skills recovered 55%–100%+ of GPT-5.4-mini's reasoning-mode advantage while using 2.7–6× fewer output tokens and zero reasoning tokens. They exceeded reasoning mode on ALFWorld (78.7% vs. 71.3%) and retail (40.8% vs. 35.0%). Skills distilled only from cheap non-reasoning trajectories remained competitive with those built from paired reasoning traces, suggesting that reusable procedural knowledge can often be recovered by wide search across episodes instead of repeatedly paying for deep search within each episode.