Check traces for hidden agent errors

Laminar’s flow-1 reviews agent traces end to end so quiet mistakes can be caught without sampling.

Done with Custom agent· Intermediate· Oct 2026

What they did
Laminar’s flow-1 is a small open-source model built for trace investigation. It was refined with supervised tuning and RL on synthetic investigation data, then run in Signals, where it starts from a compressed trace plus the user’s request, scans spans like files in a repository, opens the useful ones, and returns JSON findings in a user-chosen schema.
How it went
Laminar says it performs at GPT-6-sol level on its in-house benchmark while running at 23x lower cost, which would make full-trace analysis feasible. Similar issues are grouped together, though the benchmark is the vendor’s own.
Tips
Use a concrete review goal so the agent knows what counts as a problem, such as verifying that a browser task actually completed checkout. If you want to turn repeated mistakes into eval data, keep the clustered outputs for later review.

Try it yourself with Claude Code

Review the exported agent traces in [path to trace files]. Go through each trace and look for subtle failures, including missed steps, bad clicks, incorrect citations, or tool results that don’t line up with the final answer. Begin with the compressed trace, then open and inspect any spans that look relevant. Return JSON in [your schema] for every trace you flag, with the trace ID, the failure type, and a short note. Also list the repeating failure patterns and how often each appears. Before any irreversible action, stop and ask me to approve the traces you plan to flag and the output format.

Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.

Read the original on news.ycombinator.com ↗

Get 3 agent uses like this every morning

With the day's top AI news. Free, and one click to unsubscribe.