HERA: Harness-Environment Co-Evolution for Reliable Agentic Abstention
Han Luo, Bingbing Wen and Lucy Lu Wang (University of Washington) with Zora Zhiruo Wang (CMU) and Pan Lu (Stanford) introduce HERA, which co-evolves an agent harness and its training environments so agents learn to stop when a tool-use task is infeasible.
Ask this paper
hera-bench. A mutation pipeline builds matched feasible and infeasible versions of executable tool-use tasks; the benchmark has 60 verified pairs where the agent must act on one and abstain on the other.
Co-evolution. Failed episodes drive both new task mutations and harness edits, and ablations show both updates and multiple iterations are needed. On the final 50-pair training pool HERA reaches 86.0% abstention and 78.0% paired success, 10.0 points above the next baseline.
Held-out results. The final harness raises abstention success from 61.7% to 83.3% and paired success from 48.3% to 70.0%, while feasible-task completion also rises from 68.3% to 76.7%.
Transfer. Applied unchanged to 19 other LLMs, the harness improves abstention by 15.3 points and paired success by 12.5 points on average; GPT-5.6-Luna with HERA matches base GPT-6-Astra at about 85% lower cost.
Abstract
Large language model (LLM) agents are increasingly capable of acting in complex tool-use environments, yet they often fail to recognize when tasks are infeasible and no valid solution exists. Recent work has formalized this reliability gap as the problem of agentic abstention, and existing approaches typically optimize a model or agent harness against a fixed set of tasks, leading to limited generalization to unseen failure modes. We introduce HERA, a framework for harness-environment co-evolution for agentic abstention. HERA consists of (i) a pipeline to automatically construct verifiable pairs of feasible and infeasible tasks by applying controlled environment mutations that transform solvable tasks into cases requiring abstention, and (ii) a co-evolution procedure in which performance failures on previous tasks are used to drive harness adaptation and generate new execution environments and tasks geared towards previous weaknesses. On held-out evaluation tasks, an evolved harness from HERA improves abstention accuracy from 61.7% to 83.3% while improving feasible-task completion from 68.3% to 76.7%, achieving the highest abstention and feasible-task completion among the compared methods. The resulting best harness transfers across 19 other LLMs, improving abstention accuracy by 15.3 percentage points on average without any model-specific optimization, and enabling smaller models to match the performance of more powerful models at an estimated 85% lower cost.