It is a preview with a narrow eligibility list, and one setting deserves attention before anyone clicks Save. The details follow.
What GitHub announced
According to GitHub's changelog, administrators who connect GitHub to Agent 365 can:
- See activity from several Copilot clients: cloud agents on GitHub.com, Copilot CLI, Visual Studio Code, Visual Studio, and partner IDEs including JetBrains and Eclipse.
- Review user sessions, including prompts, responses and tool calls, and view usage trends in Agent 365.
- See more risk information in Microsoft Purview and Microsoft Defender to investigate and reduce risk.
The changelog says the feature is for GitHub Enterprise Cloud customers with Enterprise Managed Users (EMU). GitHub's setup documentation adds detail. The streaming feature is in public preview and works for enterprises that use EMU on GitHub.com or GitHub Enterprise Cloud with data residency.
The documentation also states that Microsoft Agent 365 supports Copilot agent session events only. Your full GitHub audit log, including Git events and membership or permission changes, does not go to Agent 365. Other destinations such as Splunk, Azure Event Hubs or Amazon S3 still handle that.
In short: Copilot agent session data now goes into Microsoft's agent governance console. It is still in preview, it requires EMU or data residency, and only agent session events are sent.
How this builds on earlier Copilot session streaming
Agent session streaming already existed before Agent 365 was added as a destination. In July, GitHub opened a public preview that let eligible organizations stream or retrieve Copilot agent session records across supported clients. At that point, Microsoft Purview was listed as a supported streaming endpoint in public preview.
Both Purview and Agent 365 now appear as Microsoft destinations in GitHub's documentation, and each accepts Copilot agent session events only. The change is not that the telemetry exists. The change is that the telemetry can now be sent to Agent 365, where it sits next to the other agents a Microsoft-based organization tracks, instead of a separate pipeline built only for GitHub.
The default view inside GitHub also helps explain the integration. GitHub's documentation says the audit log retains events for the last 180 days, and GitHub recommends streaming the audit log to a Security Information and Event Management (SIEM) platform, where you can view long-term history and set up alerts for anomalous activity. An analysis by the AI-tooling site Failproof AI notes that the standard actor:Copilot audit view covers cloud agent activity on GitHub, not local CLI or IDE sessions. The session stream is meant to fill that gap by including the CLI and IDE clients.
In short: Purview already received this data. Agent 365 is a second Microsoft destination, and the stream reaches local client activity that the regular audit log does not.
Setup, step by step
This is the sequence in GitHub's current documentation. An enterprise owner has to do it.
Step 1: Turn on Copilot agent session streaming
- Open your enterprise, for example from the Enterprises page on GitHub.com.
- Go to AI controls, then Copilot.
- In the Billing & usage section, enable the Copilot Usage Records Streaming policy.
The Copilot Usage Records API policy is a separate setting. GitHub's current setup page describes it as the way to retrieve usage data through the REST API, not a requirement for the Agent 365 stream. Coverage of the July preview gave different advice. ITTrip and CyberOGZ both described choosing Enabled everywhere, and CyberOGZ reported that the July changelog said both Copilot Usage Records Streaming and the Copilot Usage Records API must be enabled everywhere. If the stream stays empty, check this first. Follow the current GitHub documentation over older summaries.
Step 2: Configure the Agent 365 destination
- In your enterprise Settings, go to Audit log, then Log streaming.
- Open the Configure stream dropdown and select Microsoft Agent 365.
- Click Authorize with Microsoft Entra.
- Sign in with a Microsoft Entra account for your tenant and authorize the GitHub app.
- Choose how much Copilot activity data to share with this connection. The next section covers this choice.
Step 3: Confirm the stream is healthy
GitHub's general streaming documentation says a health check runs on each stream every 24 hours. If a stream is set up incorrectly, enterprise owners get an email. A misconfigured stream has to be fixed within six days, or audit events are dropped. Keep the owner mailbox monitored rather than letting it collect unread mail.
If you have data residency requirements, GitHub says to configure the destination to store data in the right region before you turn streaming on.
The "Include raw content" setting
This is the most important decision in the setup.
GitHub says Include raw content is selected by default. When it is on, the Agent 365 connection receives prompts, responses, tool inputs and results, and client addresses. When it is off, only activity metadata is sent, such as dates, model names, token counts and tool names.
The setting applies only to the Agent 365 connection. Your other streams are not affected.
Why this matters (my analysis):
- Prompts often contain sensitive material. Developers paste stack traces, configuration snippets, sometimes credentials, and source code into Copilot. Sending that raw content to Agent 365 creates another copy of it, which needs its own access controls and retention rules.
- Investigators need the content. Metadata can show that someone ran an agent session that called a shell tool 400 times. Only the raw content shows what those calls did.
- Client addresses can identify people. In regions with strict employee-monitoring rules, works councils and privacy teams will want to review this before it is turned on.
A practical approach is to decide on purpose, with security, legal and engineering leadership involved, instead of accepting the default. Organizations that want user attribution without collecting content have another option. GitHub's separate OpenTelemetry monitoring path includes an identity capture class that attributes agent activity to a user without collecting prompts, responses or tool content. That is a different pipeline to an OTel-compatible backend, not a toggle in the Agent 365 connector.
In short: the default sends full content. Make that choice deliberately.
What happens on the Microsoft side
Microsoft's Purview documentation for Agent 365 says audit support covers agent-to-human, human-to-agent, agent-to-tools and agent-to-agent interactions. Those events appear in DSPM activity explorer and can be searched in Purview Audit. Microsoft also describes an AI observability view under DSPM that ranks agents by risk level, using risk levels from Insider Risk Management.
GitHub's announcement does not say which Defender detections, Purview risk fields, licenses or retention periods apply to GitHub-originated events. Microsoft's general Agent 365 documentation shows what the platform can do. It does not confirm that every capability works the same way for Copilot sessions coming from GitHub. Test with your own tenant before you promise anyone a dashboard.
Operational notes for admins
From GitHub's audit streaming documentation:
| Behavior | What it means for you |
|---|---|
| Streams export events from the time streaming is enabled onward | Activity from before you turned it on is not sent |
| Files are compressed JSON with at-least-once delivery | Expect occasional duplicates in downstream analysis |
| Streams can be paused, with logs held for up to seven days | Pause for maintenance, but don't leave it paused for long |
| A daily health check emails enterprise owners if a stream is misconfigured | Fix misconfigurations within six days or events are dropped |
| Multiple endpoints are supported (also in preview) | You can send to Agent 365 and keep your existing SIEM stream |
Who benefits, and the caveats
Microsoft-centric enterprises benefit most. Organizations already using Entra, Purview and Defender can put developer AI agents under the same governance as Copilot Studio agents and other workloads, which matters as AI agents spread across a company.
The caveats:
- This is a preview. Behavior, schemas and setting names can change.
- Eligibility is limited. Standard GitHub Enterprise Cloud accounts that don't use EMU, and therefore lack centrally managed identities, cannot use it.
- Only agent session events are sent. For full audit coverage, keep your existing SIEM stream.
- The raw-content default puts the burden on admins to set privacy guardrails.
There is also a broader point. GitHub and Microsoft benefit when your agent governance lives in Microsoft's tools. That does not make the feature less useful, but teams running Splunk, Datadog or an OpenTelemetry stack can still get similar Copilot session data through those routes.
Bottom line
For eligible enterprises, this is the simplest way so far to put GitHub Copilot agent activity from Visual Studio, VS Code, JetBrains, Eclipse, the CLI and GitHub.com cloud agents into Microsoft's agent governance console. Setup takes a few clicks and an Entra sign-in. The decision that needs real thought is whether to keep Include raw content turned on.
References
- Monitor GitHub Copilot activity through Microsoft Agent 365 GitHub Changelog · 2026-11-17T18:00:59+00:00
- Monitoring agentic activity in your enterprise - GitHub Enterprise Cloud Docs docs.github.com
- GitHub opens Copilot agent session streaming preview for enterprise audit teams - CyberOGZ cyberogz.com