GitHub Copilot adds OpenTelemetry export for agent sessions, admins turn it on in managed-settings.json

For a security or platform team, a Copilot agent session has been a report from the user about what happened. GitHub turned on OpenTelemetry export inside the Copilot app on 22 September, and an enterprise administrator can now send those sessions to the same monitoring tool the rest of the stack uses. The setting is a telemetry property in the enterprise managed-settings.json file, and it is off until an admin turns it on.

What a trace contains

GitHub says the export covers "the flow of a session, including requests to AI models and the tools an agent uses", and that a reviewer can "review step-by-step traces of agent execution in their existing monitoring tools". The changelog names Splunk Observability Cloud in its demonstration and points at the OpenTelemetry standard for the wire format, which is what any OTel-compatible backend reads. Traces are the data type; the announcement does not promise metrics or logs.

How the setting works

The telemetry property lives in the enterprise managed-settings.json file that GitHub already uses for other Copilot policies, so a team that has been managing settings there does not stand up new infrastructure. The setting names the OTLP endpoint traces go to, and applies across every team the enterprise covers. Individual developers do not each configure their own exporter, and cannot opt out.

The default keeps prompts out of the trace

"Prompt and response content is excluded by default", the changelog says, and adds "review your content-capture settings before enabling it". The default trace therefore records which model was called, which tools ran and in what order, and how long each step took, without the text of the conversation. A team that wants prompts as well has to change that setting knowingly. That distinction matters, because a trace with prompts is subject to data-handling rules the exporter and the destination may not meet.

Why an audit trail is the point

The reason to route this into a monitoring tool at all is that a Copilot agent is a program that ran on a developer's behalf, called external services and edited files, and until now the record of that lived in Copilot's own storage. Sending the trace to the enterprise's SIEM or observability platform lets the incident and security teams treat an agent session like any other production trace: alert on shell tool calls, count how long a session ran, look at a session by its ID after the fact, and reconstruct what an agent did on the day something broke.

The important part of this release is what the default excludes. A trace of tool calls is safe to send off to a monitoring vendor; a trace of prompt content is a copy of every question a developer asked, and every file the agent read, arriving at a third-party endpoint. GitHub choosing exclusion as the default is the piece that makes this usable without a review.

Source

• GitHub, OpenTelemetry in the GitHub Copilot app, 22 September 2026.