HERA: Harness-Environment Co-Evolution for Reliable Agentic Abstention

HERA. Harness-Environment Co-Evolution for Reliable Agentic Abstention. By Han Luo and five others, from the University of Washington, the University of Leeds, and three other institutions. Published on the fifth of October. Its code repository is public and real: fifteen source files.

Language model agents are getting better at acting in complex tool-use environments. What they often fail to do is recognise when a task is infeasible, when no valid solution exists and the right move is to stop. Recent work has formalized this reliability gap as agentic abstention. Existing approaches tend to optimize a model, or the agent harness around it, against a fixed set of tasks. A fixed set only contains the failure modes someone thought to include, and the result is limited generalization to failure modes that were never seen.

HERA treats the task set as something that should move too. It is a framework for co-evolution: the harness and the environment evolve together. As the harness improves, the environments it is tested in change to target where it is still weak.

There are two parts. The first is a pipeline that automatically constructs verifiable pairs of tasks, one feasible and one infeasible. It does this by applying controlled mutations to the environment, which transform a solvable task into a case where abstention is the correct answer. The second part is the co-evolution procedure itself. Performance failures on previous tasks are used in two ways: they drive adaptation of the harness, and they drive the generation of new execution environments and tasks aimed at those same weaknesses.

On held-out evaluation tasks, an evolved harness from HERA raises abstention accuracy from sixty-one point seven percent to eighty-three point three percent. It does this without becoming timid: feasible-task completion also rises, from sixty-eight point three percent to seventy-six point seven percent. Both figures are the highest among the compared methods.

The best harness also travels. Applied to nineteen other language models, it improves abstention accuracy by fifteen point three percentage points on average, with no model-specific optimization. It also lets smaller models match the performance of more powerful ones, at an estimated eighty-five percent lower cost.

Knowing when to refuse a task and being able to finish a task usually pull against each other, and in these results both numbers went up together. The method's answer to unseen failure modes is not a bigger fixed set of hard cases but a generator of them, steered by the harness's own mistakes, and the improvement it buys lives in the harness, which is why it carries over to models it was never tuned on.