Skip to main content
A friction is a specific point where an agent got stuck or went wrong. Use the Frictions tab when you want to know what to fix, without reading every step of every run. Frictions come from Insights, which analyze saved traces and output. Insights are separate from rubric Evals. See Evals and Insights.

​
What counts as a friction

Examples:
  • a command or tool call that failed and was retried several times
  • context the agent could not find
  • an error the agent could not recover from
  • a long detour that did not move the task forward
  • token waste from repeated failed approaches

​
How to use them

  1. Turn on Insights at launch, or trigger Insights later.
  2. Open the experiment Frictions tab.
  3. Start from the list. Open Traces at each friction.
  4. Fix the interface or product.
  5. Launch again and watch pass rate rise.
You can re-trigger Insights later without redoing Evals. If agents keep failing at the same step, that step is what a skill, MCP server, CLI, or SDK on your treatment should address. Keep the task’s files and machine fixed so you can trust the comparison. See Core concepts. Ask Ask Oqo to summarize the top frictions and suggest what to change next.