Microsoft's CASD: a coding agent reading full logs beats GEPA at prompt optimization by 5.7 points
dair_ai · x · 2026-09-25
A new Microsoft paper on prompt optimization introduces CASD, arguing that instead of running search loops over small batches of trajectories, you should hand a coding agent your full set of agent logs and let it write the analysis code.
Key points:
- CASD uses an off-the-shelf coding agent to compute statistics over the entire trajectory corpus, find recurring failure modes, read representative episodes, and write the findings as rules in one prompt
- Requires no environment access and no validation data
- Across ALFWorld, tau2-bench (retail/telecom), and Spreadsheet Bench-Verified, one pass improves the unoptimized baseline by 16.6 points on average, vs GEPA's 10.9 and SkillOpt's 5.3
- Each optimized prompt costs about $1.60, over 22x cheaper than validation-gated search
More from coding & agent
- Micius-Flash: 27B Multimodal Model for Hardware and Semiconductor Agents — zhiman-ai · 2026-11-11
- Jane Street's AI Skill Is a Strict Socratic Tutor You Can Copy Today — omojumiller · 2026-11-11
- Designer builds 100k-site German heritage map with Claude, zero code written by hand — neutronenkind · 2026-11-11
- Matt Pocock: DDD, not agentic tricks, is what blows minds in his AI course — mattpocockuk · 2026-11-11
- AI agents run 24/7 for a month to rewrite Météo-France's AROME weather model — IgorCarron · 2026-11-11
- Prompting secret: ask the model to roleplay a legendary high-agency explorer — tokenbender · 2026-11-11