Skip to main content
Open Agent Picks in your project. Check the test date and source before comparing numbers. Public test means category results are being shown because your project has no finished test of its own. Re-test comparisons live with their Actions and in Results.

​
Overview

Overview summarizes which tools were installed, your tool’s install rate, what it lost most often to, and results by framework. Use the Agent and Question filters to narrow the measurement. The install-rate trend and test history compare paid tests over time. Unfinished tests have no rate. If a test did not recognize your tool, its missing rate is not a zero. Select Load earlier tests to extend the history.

​
Matrix

Matrix shows the environments as rows and the requests as columns for the selected agent. Select a cell to see the tools installed there, the completed run count and likely ranges. Use it to locate a specific gap, such as a framework where agents build their own solution when asked for an integration. Check the cell’s coverage before treating its rate as representative.

​
Runs

Open Runs to browse your project’s starter and paid run history. Select a run to inspect its agent, model, environment, status, recorded pick and retained evidence. The trail can contain search, fetch, file-read, command, install and file-write observations. Run details withhold request and job text, search terms, file paths, raw commands, error text and artifacts. URLs show only their origin, and package names are limited to recognized tools. A withheld field does not mean the activity never happened. If no evidence events are retained, the run may not have recorded them yet. Public or copied category results do not give access to the source project’s individual runs. Observations and HTTP or command statuses do not prove why an agent chose a tool or that its integration worked.

​
Install rates and likely ranges

An install rate is the number of completed runs that picked a tool divided by the completed runs in the selection. For example, four installs in eight completed runs is 50%. Each rate includes its run count and a Wilson 95% likely range. Small samples have wide ranges. If two ranges overlap, the measurements do not establish a clear difference. Keep the range beside the rate when sharing a result. These outcomes remain separate:
  • Built its own means the agent implemented the capability itself instead of choosing a named tool.
  • No provider chosen means it wired in a hosted service without naming a provider.
  • Another tool means a job run added a package or service outside the confirmed tools.
  • Not classified means the evidence did not establish a choice. It is not proof that the agent chose no tool.
If your tool was absent from the tools recognized by the test, it is not measured, not a zero-install result.

​
Coverage

Rates use completed runs only. Always read completed / planned alongside them. Failed runs and runs that did not start leave gaps; they do not count as completed losses. A cell with no completed runs is Not measured. Partial coverage can hide differences between environments, requests or models even when an overall rate is available. Use Losses, Actions and re-tests to turn measured gaps into fixes and compare the next measurement.