Can AI Agents Identify What State Their Actions Depend On?

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Long-running AI agents may act on state observed earlier in their execution, creating a risk that relevant information changes before an action is performed. Revalidating the entire observed state before every action is costly and unnecessary; instead, an agent should revalidate only the facts on which its action depends. We formulate action dependency identification: given a user task, observed state represented as atomic facts, and a proposed action, identify the facts that should be revalidated before execution. We construct a benchmark derived from WebArena-Verified tasks and evaluate models across capability levels and reasoning configurations. High-capability models achieve up to 0.99 F1, while controlled comparisons show that inference-time reasoning improves dependency identification at substantial latency cost. Error analysis further shows that categorical and relational state are more frequently missed than numeric, temporal, and attribute state.