Agents spot a blocked task but keep going: writing feasibility into the harness lifts strict self-awareness from 20% to 60%

The evaluation splits agent self-awareness into three parts: judging whether a task is feasible, acting on that judgment, and holding or revising it under challenge, while separating what a task permits from what the agent can solve. Earlier success-oriented evaluations only checked whether a task was completed, so they could not separate recognizing a blocker and continuing anyway from never recognizing it. In a Terminal-Bench 2.1 protocol baseline, the agent identifies the true blocker on a high share of blocked search tasks, but strict self-awareness reaches only 20%; the gap comes from gates against fabrication, from requiring the task agent itself to author the abandon action, and from requiring the correct reason.

Harness-level search improves this behavior without updating model weights, treating prompts, processors, control flow, and stopping behavior as searchable objects. Previously this depended on model capability or hand-edited prompts, with no comparison that put feasibility judgments directly into the behavioral objective. Abstention-only search raises strict self-awareness on blocked Terminal-Bench search tasks from 20% to 60% and in retail from 20% to 40%; a two-sided search that also rewards solving feasible tasks reaches 60% strict self-awareness and lifts feasible solve from 47% to 60% in Terminal-Bench, and 44% and 47% respectively in retail.

The recognition-action gap becomes measurable and yields a two-sided target: stop when a task is infeasible, and keep working when it is feasible. Optimizing abstention alone invites the opposite shortcut of refusing work that is actually feasible. On retail's held-out set, the abstention-only harness stays at 25% strict self-awareness, matching its initial reference, while feasible solve falls from 41% to 34%; the joint-selected harness records 34% strict self-awareness and 47% feasible solve, with zero false abstentions on the 32 feasible tasks.

The evaluation targets two controlled agent settings, terminal and interactive, and uses paired feasible and blocked tasks to measure both stopping when warranted and continuing when not. It suits readers studying the consistency between an agent's feasibility judgments and its behavior, and engineering teams that need a two-sided objective for an agent harness. The task artifacts, validation procedures, and scoring rules are designed to be reused in other agent settings.

The largest harness gains appear on the search tasks used to select the harnesses, and held-out gains remain limited; retail held-out cells contain one stochastic trajectory per task, so two extra strict self-awareness successes do not establish a reliable improvement. Terminal-Bench blocked held-out tasks have no full strict self-awareness evaluation, so whether search-set gains transfer to new blockers is unresolved. Retail's semantic scoring uses a judge built on the same model as the task agent, which may share model-specific errors; feasibility probabilities come from a separate read-only probe and need not reflect what the agent believed while performing the task. The stability test injects one neutral challenge and does not show behavior under repeated or adversarial pushback.