The four questions
Each analysis feature answers a different question about production behavior.Patterns
Is behavior we already care about happening? Reusable classifiers for known behaviors: frustration, confusion, task failure, or any custom signal you define.
Failures
What recurring problems keep happening? Errors from runs and tool calls, grouped and ranked by frequency and recency.
Insight Search
Why is this happening? Natural-language investigation across runs, sessions, tools, models, costs, and feedback.
Recaps
What changed that I should pay attention to? AI-generated daily, weekly, or monthly reports with linked evidence.
Dimensions you can slice by
Every analysis works across the same first-class dimensions, because they are stored as native columns rather than free-form metadata.Conversations and sessions
Runs grouped into full user conversations.
Users and organizations
End users and the customer accounts they belong to.
Tools and models
Which tools and models were involved, with cost, latency, and error breakdowns.
Feedback
Thumbs up / down and text feedback per run.
A typical workflow
1
Something surfaces
A Recap flags a spike, a pattern crosses its alert threshold in Slack, or you notice a change on the overview dashboard.
2
Group the affected runs
Open the pattern or failure. Both list the runs (and sessions) that match, ranked by recency and frequency, with links into each trace.
3
Investigate root cause
Open a representative trace. Inspect the model steps, tool calls, arguments, results, and errors that led to the outcome.
4
Quantify impact
Filter by user, organization, tool, or model to see how many customers are affected and where the problem concentrates.
5
Ship the fix and monitor
Change the prompt, tool, or workflow. Create a custom pattern to track the behavior going forward, and let Recaps and Slack alerts flag regressions.
Related
Traces
The primary evidence behind every analysis result.
Silent tool failure walkthrough
Discovery → trace → root cause → monitoring, end to end.