Releases

The Releases tab of an application shows how each release of your agent behaved, from its traces of the last 7 days, in the environment you chose.

A release is the commit a trace says it came from (vcs.ref.head.revision), or its service.version when it names no commit. Traces that name neither are not grouped. See Name the code version.

The release table

Column	What it shows
Release	The commit (first 10 characters) or the version
Runs	How many traces it sent
Failed	The share of traces with a failed step
p50, p95	The median and 95th-percentile duration of a trace
Tokens / run	The tokens a trace used, at the median
Last seen	When its newest trace arrived

Compare two releases

Select Compare on a row to compare it with the release before it, or choose any two under Compare … with …. The comparison shows:

• runs, failure rate, p50, p95 and tokens per run, before and after, with the change;
• Steps whose failures rose: each step (a tool or another span name) whose failure rate went up, by how many points, with Open trace for an example.

A rise is evidence about what changed in your agent’s behaviour, not proof of a cause. Open the example trace and Find in code to check.

Run evaluations on a release

Run evaluations opens Evaluations with that release’s commit chosen, so the application’s test set runs on exactly that code in a fresh sandbox. It needs the application’s GitHub repository, because the run checks out that commit. Without a repository, runs on recorded traces still work from Evaluations.

Over MCP

A coding tool reads releases with releases_get, alone or two side by side. See Tools and prompts reference.