Investigating what an agent did does not necessarily require retaining everything it read. GitHub Copilot's October 12 update separates identity from content in OpenTelemetry: an enterprise can attribute an execution to a user without collecting prompts, responses or tool results.

The update extends September's support for exporting execution traces to an observability backend. A trace connects session steps; token and duration metrics help locate bottlenecks. The choice of how much to observe is now more granular, and so is the responsibility for that collection.

Identity and content are separate decisions

The telemetry.capture.identity control can add the GitHub login, operating-system user and machine name. It is disabled by default and does not change the existing pseudonymous identifier.

Five content classes can be enabled independently:

Control

prompts

What collection can include
Prompts and supported input context
Control

responses

What collection can include
Responses and supported reasoning content
Control

toolArguments

What collection can include
Arguments, commands and tool identifiers
Control

toolOutput

What collection can include
Tool results and error details
Control

policyDetail

What collection can include
Raw policy details, such as sandbox paths

All default to off. Legacy captureContent remains a shorthand for those five classes, but does not enable identity. In the clients documented for managed configuration, explicit enterprise values override local settings.

That separates two easily confused needs: attributing a session during an investigation and reconstructing its content. The latter requires an additional decision. Identifying someone is unnecessary for every cost analysis; retaining their code is unnecessary for discovering that a tool takes too long.

The client changes the control boundary

Full granular capture is documented for Copilot CLI, the Copilot app and VS Code sessions using Agent Host. VS Code's local harness supports identity and legacy content capture, but does not apply the five classes independently.

This rules out a dangerous shortcut: publishing one configuration and assuming every client interprets it identically. VS Code's documentation also describes extension-specific precedence rules, including environment variables. Before standardizing collection, verify the team's client, version and execution mode.

Telemetry configuration identifies the receiving collector. The observability service therefore joins the project's trust boundary. Permissions, retention and trace access should reflect content sensitivity, particularly when capture includes files, commands or directory paths.

A trace can explain an agent's failure. It can also create another copy of data the team needed to protect.

Observability still cannot measure quality alone

Consider a session with few tokens and low latency. It may have been efficient, or it may have skipped the test that would reveal a regression. A more expensive session may have completed difficult work with less rework.

The useful correlation connects traces with reviewable results, executed tests and corrections after review. Identity can help investigate a specific incident, but an individual productivity ranking based only on tokens would confuse activity with delivery.

Start with an operational question, such as finding repeated calls or explaining tool failures. Collect the minimum data required and expand capture only for a defined reason. Enterprise policy controls help distribute rules; the quality of those rules depends on who decides what should be collected.

Sources: initial OpenTelemetry announcement, granular capture and identity, October 12, managed settings reference and VS Code agent monitoring.