Microsoft researchers introduced Coding-Agent Skill Distillation (CASD), a method for creating optimized system prompts by analyzing existing agent logs. The approach processes a full corpus of past task trajectories to identify behavioral patterns and failure modes, rather than relying on iterative search or live environment interaction.
According to the paper, CASD costs approximately $1.60 per pass, which is over 22 times cheaper than standard validation-gated iterative tuning methods. The technique does not require held-out validation sets, reducing both preparation time and the risks associated with testing prompts against live production systems.
In testing, CASD achieved an average performance gain of 16.6 percentage points over unoptimized baselines. It outperformed the widely used GEPA optimizer on three out of four benchmarks, even when competitors were provided with additional validation data and unrestricted access to environments.
GEPA remains a standard tool at companies such as Databricks, Shopify, Dropbox, and OpenAI. While CASD proved superior in most cases, the researchers noted that iterative methods like GEPA still outperformed it on specific structured tasks, such as those measured by SpreadsheetBench-Verified.
