Harness optimization—refining an agent’s prompts, tools, and workflows—has become a powerful way to boost agents built on frozen or closed‑source large language models (LLMs). Existing approaches such as the Darwin‑Godel Machine and the Huxley‑Godel Machine (HGM) rely on recursive self‑improvement and were primarily designed for coding, where the competence needed to improve the agent aligns with the competence being improved.
Planning tasks, however, require an agent to produce a sequence of tool calls that jointly satisfy interacting constraints. This misalignment causes open‑ended self‑modification to repeatedly explore ineffective changes, consuming a large evaluation budget.
To address this, we introduce HGM‑Dual, a diagnostic‑guided framework for budget‑efficient harness evolution. First, HGM‑Dual represents agent evolution as lineage‑aware context, recording what each ancestor changed, which behaviors improved or regressed, and which modifications were ineffective. This gives the editor explicit edit‑outcome attribution instead of an unstructured archive of past agents.
Second, HGM‑Dual employs error‑conditioned dual‑path child generation. For each selected parent, it creates a general child that preserves open‑ended exploration and specialized children that target the harness components most likely responsible for the parent’s highest‑ranked constraint violations. Experiments show that HGM‑Dual outperforms the strongest baseline by up to 14.45% relative improvement in case accuracy while matching HGM variants’ final performance with 2‑4× less evaluation budget.
Review