Back to AI Research

AI Research

HGM-Dual uses failure history to guide planning-agent harness changes

Key Takeaways

  • The framework records edit outcomes and targets constraint violations, reporting more efficient search than open-ended harness evolution in its experiments.
  • HGM-Dual studies how to improve an agent’s prompts, tools and workflow without changing the underlying language model.
  • The paper focuses on planning tasks, where a sequence of tool calls must satisfy interacting constraints.
  • It argues that successful self-editing methods from coding do not automatically transfer to this setting.
  • The [HGM-Dual paper](https://arxiv.org/abs/2610.30042) represents agent evolution as lineage-aware context.

HGM-Dual studies how to improve an agent’s prompts, tools and workflow without changing the underlying language model. The paper focuses on planning tasks, where a sequence of tool calls must satisfy interacting constraints. It argues that successful self-editing methods from coding do not automatically transfer to this setting.

Preserve the history of changes

The HGM-Dual paper represents agent evolution as lineage-aware context. It records what an ancestor changed, which behaviors improved or regressed and which edits were ineffective.
That turns a collection of past agents into a more explicit account of edit outcomes. The intended benefit is to help the next editing step avoid repeating unsuccessful changes and connect a proposed modification to a particular observed problem.

Keep exploration and targeted repair

For a selected parent, HGM-Dual creates a general child for open-ended exploration and specialized children aimed at harness components associated with the highest-ranked constraint violations. The paper calls this error-conditioned dual-path generation.
The two paths address a practical tradeoff. Targeting a known failure may use an evaluation budget more efficiently, while unrestricted exploration can still uncover improvements the diagnosis did not anticipate. The framework does not assume that every failure has a single obvious component-level cause.

What the reported efficiency means

The authors report up to 14.45 percent relative improvement in case accuracy over their strongest baseline. They also report matching HGM variants’ final performance with two to four times fewer evaluations. Those are results within the paper’s planning experiments, rather than a universal reduction in the cost of improving any agent.
For teams studying harness optimization, the useful design idea is to preserve the relationship between a change and its observed consequences. An evaluation should still check for regressions on held-out tasks and count the full search budget. A convincing improvement needs both a better final agent and an honest account of the experiments used to find it.

Comments