An agent changes several files, calls tools, talks to a model and delivers a patch. If the result is surprising, where do you start investigating? Many teams can see the final output and aggregate consumption, but have less visibility into the path taken during an individual session.
GitHub's September 22 announcement adds OpenTelemetry support to the GitHub Copilot app. Administrators can configure export through managed-settings.json and send the data to compatible observability tools. It is less flashy than a new model in the picker, but potentially more useful for teams operating agents at scale.
What becomes visible
GitHub's technical documentation describes three data types: traces, metrics and events. A trace connects steps in a session, such as a model call, a file read and another model call. Metrics include input and output token counts. An event can record whether a suggested edit was accepted or rejected.
That helps teams ask practical questions. Did an agent spend too long finding context? Did it call the same tool repeatedly? Did a short task consume an unusual number of tokens? The announcement does not promise a dashboard that answers everything automatically. It provides the raw material to investigate, correlate data and build useful indicators.
One scope limit matters: the announcement describes configuration for the Copilot app through enterprise-managed settings. Teams should not assume, without checking each client's documentation, that every environment where Copilot runs will emit identical data.
Observability is not a quality score
A trace shows the route taken; it does not prove that the delivered code is correct. A session with fewer calls might be efficient, or it might have skipped a necessary check. More tokens can signal waste or a genuinely difficult task.
The more useful view connects telemetry to outcomes: time to a reviewable change, tests run, rework after review and production failures. For teams running many agents, that correlation may reveal where costs rise without improving delivery. It is an operational hypothesis to test, not an ROI promise in GitHub's announcement.
Privacy also needs a deliberate choice. GitHub says prompts, responses and tool arguments are excluded from export by default. Content capture can be enabled, but it may include sensitive code and data. Before collecting more, teams should decide who receives traces, how long they are retained and which problem the extra collection is meant to solve.
This release brings agent observability closer to familiar software operations practices. The next move is not to measure everything. Pick two or three questions your team cannot answer confidently today, then see whether these traces actually help.