FAPO

Most prompt optimisation tools treat a pipeline as a black box. FAPO does not. Cisco Foundation AI has open-sourced Fully Automated Prompt Optimisation, a system that evaluates each step in a multi-stage LLM chain, pins down exactly where things go wrong, and proposes fixes at the prompt, parameter, and chain-structure level.

It uses Claude Code as the orchestrator. An independent reviewer validates each proposed variant before anything gets accepted. That two-stage check is the bit we find most sensible: automated optimisation without any kind of validation loop tends to drift in odd directions.

The benchmarks show FAPO outperforming GEPA on 15 of 18 model-benchmark comparisons, which is a reasonable signal that the step-level attribution approach is doing real work rather than just shuffling wording around.

If you are building anything with chained LLM calls, this looks worth a serious look. The failure attribution alone could save a lot of manual debugging.