ax-agent-optimize

To optimize an AxAgent workflow, you generate tuning and evaluation code using the @ax-llm/ax library. This involves configuring judgeOptions, eval datasets, and optimization targets to automate and refine agent behavior across multiple rounds.

Recursive optimization with AxAgent iteratively refines agent behavior by running multiple evaluation rounds. The process uses deterministic metrics and built-in judge paths, saving and loading optimizedProgram artifacts between cycles to ensure consistent, repeatable improvements.

Configuring agent.optimize evaluation datasets requires defining clear task designs and deterministic metrics within the @ax-llm/ax framework. You must structure eval datasets and judgeOptions to measure specific optimization targets accurately for the AxAgent.

You save and restore optimized agent artifacts by managing optimizedProgram files within the @ax-llm/ax agent.optimize workflow. This artifact management ensures repeatable patterns and deterministic outputs during recursive AxAgent tuning and evaluation.

Yes, you can configure built-in judgeOptions for AxAgent evaluation using the @ax-llm/ax library. This allows you to establish deterministic metrics for codegen tasks without relying on external evaluation tools.

Recursive agent optimization requires strict artifact management and deterministic patterns to avoid drift. If eval datasets or judgeOptions lack clear task design, the @ax-llm/ax optimization targets may not yield repeatable or effective codegen results.