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

HERA automates agentic abstention training by using controlled environment mutations to transform solvable tasks into infeasible ones, creating verifiable failure modes. This co-evolutionary pipeline generates paired feasible and infeasible datasets to train agents to recognize task impossibility, improving generalization to unseen failure states beyond fixed task sets.