Agentic self-awareness improves AI task decision making
Agentic Self-Awareness: Aligning Knowing, Acting, and Holding
Artificial Intelligence
Summary
AI agents often face tasks that seem solvable but can have hidden problems or blockers that show up only when trying to complete them. The authors describe "agentic self-awareness" (ASA) as the ability of an AI to know when a task is feasible, act accordingly, and reconsider its actions when new obstacles are found. They created tests to measure how well AI recognizes blockers and chooses correct responses, finding that many agents detect problems but fewer take the right action in response. This work introduces a system to better measure and improve AI decisions when faced with uncertain or changing tasks.
What this means in practice
- •For ai system engineers: Improve AI agent reliability by integrating self-awareness checks that identify task blockers and trigger correct deferral behaviors during execution.
- •For robotics developers: Enhance autonomous robots' task performance by applying ASA evaluations to prevent futile attempts on impossible or conflicting assignments.
- •For customer support platforms: Use ASA principles to build AI that recognizes when it cannot solve a problem and appropriately escalates or defers action to a human operator.$Commercial implications: Enables creating AI assistants that accurately decide when to defer issues, improving service quality and reducing errors.
Authors
Qinglin Chen, Hiroaki Hayashi, Jiaxin Zhang, Chien-Sheng Wu
Abstract
Frontier AI agents increasingly act in environments where seemingly solvable tasks may lack prerequisites, contain conflicting requirements, or reveal blockers only during execution. We formulate agentic self-awareness (ASA) as the alignment of task-grounded feasibility judgments (know), warranted behavior (act), and responses to reconsideration (hold), distinguishing what a task permits from what an agent can solve. We build paired feasible and blocked tasks in Terminal-Bench 2.1 and retail $τ^2$-bench and introduce an authorship-aware M0-M3 rubric that separates recognizing a blocker from taking the certified deferral action for the correct reason. Separate probes assess stated feasibility beliefs and responses to a single challenge. In a Terminal-Bench protocol baseline, blocker recognition reaches 84.4%, but strict self-awareness (SA) reaches only 40.6%. Harness search without model-weight updates increases strict SA from 45.3% to 93.8% on reused Terminal-Bench search tasks; selection for abstention alone, however, can reduce success on feasible tasks. Independently initialized joint-objective searches improve both blocked-task SA and feasible-task solve on their own search sets. Retail held-out gains remain limited, and Terminal-Bench blocked-task transfer is not measured. This evaluation pipeline makes the recognition-action gap measurable and sets a two-sided target for improving agent decisions under changing task evidence.