Justin, founder of the AI search engine Phind, is tackling a core debugging problem for LLM agents. He's built a tool called Trails to visualize and analyze the complex, multi-step traces from browser-use agents. This addresses the lack of explicit user feedback, making it easier to pinpoint where automated tasks fail.

The demo showcases 83 browser-use trajectories from GPT-5. Traditional debugging involves manual log review or hoping system fixes work. Trails offers a direct way to query and understand agent behavior, which is critical as tasks become more complex and traces grow longer, saving developers significant time.

Future plans include live querying for running agents and using preference models to infer success from sparse data. Justin is seeking feedback and is open to collaborating with teams generating over 10k daily traces who need better debugging tools.