Show HN: 83 browser-use trajectories, visualized

Justin, creator of Phind, is building a trace visualization and analysis tool for LLM agents to help developers understand where their agents fail. The demo showcases 83 browser-use agent trajectories with plans for live querying, preference models, and sparse signal expansion for developers with high-volume trace data.

Key Points

• •Agent trace analysis is critical for debugging LLM applications, but traces are longer and more complex than traditional search logs, making manual review inefficient
• •Build visualization tools that let developers analyze LLM agent outputs directly to identify failure points and root causes without relying on sparse user feedback
• •Implement trace querying capabilities that can retroactively search past failures from currently-running agents to identify patterns and systemic issues
• •Use preference models to expand sparse feedback signals—when <1% of users provide explicit feedback, ML models can infer quality from implicit signals
• •Target high-volume agent deployments (10k+ traces/day) where manual inspection is impossible but systematic analysis could unlock significant improvements
• •Create interactive visualization dashboards that make complex agent trajectories understandable and navigable for developers debugging agent behavior
• •Prioritize observability tooling for agent developers as a key infrastructure gap—similar to how search engines needed better failure analysis tools

Workflow Diagram

Concepts