When should an agent stop acting and admit failure?

Standard LLM abstention is a single-turn answer-or-abstain choice. Agentic Abstention reframes it as a sequential decision problem: at each turn an agent can answer, abstain, or gather more information, and — crucially — the need to abstain may only become clear after interacting with the environment. A task can look feasible at the outset and reveal itself as unresolvable only several tool calls in. Evaluating 13 LLM-as-agent systems and 2 scaffolds over 28,000+ web-shopping, terminal, and QA tasks, the finding is not that agents lack the ability to abstain but that they lack the timing: some never abstain when they should, while others abstain only after many wasteful interactions.

This isolates a competence that task-completion benchmarks systematically miss. Reliability requires not just stronger execution but better judgment about when continued action stops being useful — a failure of stopping, not of capability. It connects to Why do RL agents stop asking informative questions?: both describe agents mismanaging low-information regimes, but where self-locking is over-caution, late abstention is under-caution — the agent keeps acting past the point of usefulness. The paper's fix, CONVOLVE, is notable because it distills interaction trajectories into reusable stopping rules supplied as context, improving timing without updating parameters — placing the missing competence in the prompt/scaffold layer rather than the weights, which means stopping behavior can be repaired at inference time.