Decompose observation→action into two prompts and evolve them with environment returns, not hand-tuning

arXiv 2606.17838·medium signal

This automated prompt-optimization framework splits an agent's pipeline into a goal-conditioned descriptor agent (turning observations into useful state) and a separate action-selection agent, then refines each module's prompt through an LLM-driven evolutionary loop scored by real environment returns. Unlike generic meta-prompting against a static eval set, the optimization signal comes from grounded outcomes in the agent's actual task loop. For builders, it's a template for self-improving prompts in any environment that emits a reward or success signal, without manually editing few-shot examples.