What Did the Agent Actually Do? Evidence-Grounded Oversight for Long-Horizon Agents

What Did the Agent Actually Do? Evidence-Grounded Oversight for Long-Horizon Agents. By Zhongxiang Sun and seven others, from Renmin University of China and Kuaishou Technology. Published on the fifth of October. Its code repository is worked out: about two hundred and fifty source files, with a readme and a licence.

As agents take on long-horizon tasks, the user's role changes. Instead of making each decision, the user oversees an execution that runs on its own. That is harder than it sounds. Agents produce a large volume of activity, and the evidence that would show what they did is scattered across it. The result is that it becomes difficult to tell which of the agent's decisions actually deserve a user's verification. The authors study monitors meant to help with exactly this: identifying the consequential decisions, and locating the evidence a user needs to judge what those decisions imply.

To measure the problem they introduce AgentMonBench, a software-engineering benchmark with three subsets. Together the subsets cover two complementary dimensions. One is alignment between the stated requirements and the agent's behaviour. The other is awareness of consequential autonomous decisions that call for verification. To help monitors make these judgments, the authors propose the Evidence-Grounded Behavior Graph, or EBG, a training-free method.

EBG takes evidence that is linked back to its source and groups it into behaviours. It then organizes the relationships among those behaviours into a graph. A monitor does not read the graph whole. Instead, EBG presents task-oriented views of it, so that a monitor can interpret each behaviour in its context rather than wading through the raw record.

Experiments span eight models. Compared with giving the monitor direct access to the original context, EBG improves both decision identification and evidence localization in most settings. Further experiments show that the gains in evidence localization persist across input scales and across hyperparameter settings. Real-world applications are offered to illustrate the method's practical value for human oversight.

The question in the title is the one a user overseeing an agent keeps asking. The answer proposed here is structural: turn a sprawling trace into behaviours tied to their evidence, and show the monitor the part of that structure that bears on the task at hand.