In a software-engineering benchmark, organizing an AI agent's work and its source evidence into a graph helped eight language models identify decisions that a person may need to review. The study tests monitor outputs on benchmark tasks; it does not show whether people make better decisions while supervising live agents.
Long-running coding agents can make choices that affect the result, from how they interpret an incomplete request to whether a change preserves the intended behavior. The clues may be spread across the task description, code, tool calls and their results. AgentMonBench asks a separate monitor to decide which choices warrant a user's attention and to point to evidence for that judgment.
The authors' Evidence-Grounded Behavior Graph, or EBG, rearranges those records into task-focused views. It links evidence to the code or interaction where it came from, groups related activity into behaviors and shows explicit relationships among them. The graph builder and view preparation use deterministic rules without calling a language model; a separate model reads the view and judges whether a decision should be surfaced.
The benchmark contains three 100-example tests. In SpecGAP, a monitor receives a shortened task description and access to the code, then looks for consequential requirements left unspecified. The omitted condition and its code mapping are hidden and used only to score the answer. SilentSwap changes what code does while leaving existing tests passing. FeedbackTrace uses earlier coding-agent conversations: later user feedback identifies a decision that should have been checked, while the monitor sees only the conversation up to that point.
Across eight models from five families, EBG raised decision-identification and evidence-location scores over direct access to the original repository or trace in most tested settings, though not on every metric or model. For GPT-5.6 Sol, SpecGAP's code-localization F1 rose from 0.501 to 0.569, while FeedbackTrace's evidence-hit rate rose from 0.550 to 0.610. The first score measures how well supporting code is located; the second counts whether at least one exact evidence reference was found. They are separate benchmark measures, not overall accuracy.
Finding evidence and understanding what it means remain different tasks. The paper reports cases where evidence retrieval improved while the monitor's interpretation did not. Six semantic measures were scored by two language-model judges and averaged; three evidence-location measures used rule-based scoring.
The authors also tried the method in a Codex review harness across five research-task cases, including one where cached API responses could not establish live connectivity. Those examples show how a monitor might flag a choice for checking, but they are not a controlled study of users or a deployed oversight product.
Sources
- The authors' arXiv paper, “What Did the Agent Actually Do?” (v1, submitted Oct. 5, 2026)
- EBG code repository (MIT license)
- AgentMonBench dataset card
- Figure: EBG method diagram, Figure 4 by Zhongxiang Sun et al. from the paper, CC BY 4.0.