As AI agents move from simple question answering and tool use toward autonomous workflows that can last hours or even days, a new oversight problem is emerging. Users are no longer making each individual decision; instead, they are supervising autonomous execution, often without a clear picture of what the agent actually did along the way.
A new research paper, What Did the Agent Actually Do? Evidence-Grounded Oversight for Long-Horizon Agents, examines this gap. The authors note that giving an agent a goal does not mean every constraint and decision has been fully specified. Underspecified requirements, silent behavioral deviations, and unverified consequential decisions—such as changing evaluation metrics, filtering data, or adjusting experimental settings—can shape outcomes without user confirmation.
To study the problem, the authors introduce AgentMonBench, a software-engineering benchmark with three subsets—SpecGAP, SilentSwap, and FeedbackTrace—covering two complementary dimensions: alignment between requirements and behavior, and awareness of consequential autonomous decisions that warrant verification.
They also propose the Evidence-Grounded Behavior Graph (EBG), a training-free method that groups source-linked evidence into behaviors and organizes their relationships into a graph. EBG presents task-oriented views of this graph so oversight monitors can interpret behavior in context. The authors integrated EBG into the Codex Harness with oversight checkpoints before execution, after consequential execution changes, and before final reporting.
Experiments across eight models show that EBG improves consequential decision identification and evidence localization in most settings compared with direct access to the original context. Further experiments indicate that its evidence-localization gains persist across input scales and hyperparameter settings, and case studies from real research workflows illustrate its practical value for human oversight.
The authors frame the work as a shift from model-level interpretability toward long-horizon behavioral oversight—from Human-in-the-Loop participation toward meaningful human oversight at the moments that matter.