Skip to main content
While Evals grade whether a run passed, Insights examine how the agent worked—surfacing failed tool calls, missing context, detours, and wasted retries—and group them into your project’s frictions. Two models collaborate on Insights: the Insights model analyzes each individual run, and the Experiment insights model clusters findings into project-wide frictions and writes the Brief on Comparisons. Insights do not inspect changed files or alter Evals verdicts.

Turn on Insights

1

Assign Insights models

Set your Insights model and Experiment insights model under Organization settings → Oqoqo AI defaults.
2

Enable Insights on an experiment

Keep Insights toggled on when you launch, or select Run insights on a finished experiment.
3

Review frictions

View grouped problems on the experiment’s Frictions tab, All frictions, and inline on each run’s Traces.

Choose a mode

By default, Insights run in Standard mode (one fast pass per run). If your organization has enabled Agent mode (private beta), you can switch to Agent at launch or when rerunning so Oqoqo’s insights agent reads the full run and adds What the runs did differently to Comparisons. See Agent mode for trade-offs and model rules.

What Insights read

Your Insights model reads each run’s recorded trajectory steps. If a run exceeds the model’s context window in Standard mode, a note above the Timeline on Traces (and in Compare) reports any summarized steps, shortened long steps, truncated friction signals, or traces over the 32 MB limit. No coverage note means the model read the entire run (older runs before coverage tracking show no note). If a run fails to fit even after condensing, assign an Insights model or Experiment insights model with a larger context window and rerun Insights.

Which models ran

Each Insights attempt records the Insights model and Experiment insights model used at runtime. You can inspect them above a run’s Timeline, via the help icon beside Frictions, on the experiment’s Frictions tab, in the experiment info popover, in CLI and plugin attempt reads, and in the insights and frictions datasets of oqoqo data query.

Pending frictions

Frictions populate as runs finish and cluster:
  • While Insights are still clustering, the experiment’s Frictions tab shows Warming up, and per-run frictions appear on individual run Traces until grouping completes.
  • Rerun Insights replaces each run’s previous frictions while preserving stable frc_ IDs across experiments.
  • In the CLI and plugin, oqoqo results analyze and oqoqo frictions list report coverage counts so you know whether an empty list is final or still pending.

Run Insights again

Rerun Insights anytime to re-analyze existing trajectories—for example, after switching to a newer Insights model—without spending agent runs. Oqoqo regenerates the experiment Brief when Insights finish.
Open the experiment and select Rerun Insights (or Run insights), optionally overriding the Insights model, Experiment insights model, or mode.

From the CLI and the plugin

Next steps

Frictions

Rank problems by unrecovered runs and see where to fix them.

Agent mode

Enable deep run-pair comparisons and verified frictions.

Run details

Jump straight to the trace steps cited by frictions.

oqoqo evals and insights

Full CLI reference for Insights commands.