XAgent finds bugs by running code, not just reading them

A new coding agent called XAgent scores 62.0% on SWE-bench-lite by watching code run instead of just reading the bug report first.

Most bug-fixing agents start with the issue text, then go hunting through the codebase for the function to change. XAgent's authors argue that approach has a built-in flaw. The issue description narrows the model's attention to whatever the reporter happened to mention, and misses whatever the reporter didn't. Their paper adds a step that runs the code and watches what actually happens during execution, before deciding what to fix.

The number that matters here isn't the headline score. It's function-localization accuracy: 72.8%, using dynamic behavioral analysis instead of static text matching. That step decides whether the fix goes into the right file at all. Get localization wrong and nothing downstream saves the patch.

The paper also reports XAgent resolving 7 issues that top baselines missed entirely, while staying cost-competitive on inference. No claim that it's cheaper, just that the added execution step doesn't blow up the budget.

This lines up with something I keep seeing in agent research this year, from a different angle each time: the runtime knows things the prompt doesn't. A stack trace, a variable's actual value at the point of failure, a function that gets called but never mentioned in the ticket. Text-only agents are working from a summary. Agents that execute are working from the source.

Worth watching whether this becomes standard practice or stays a research pattern. Running code during triage costs more compute and more engineering to sandbox safely, and plenty of production agent setups skip it precisely because static analysis is cheaper to ship.