M365 Agent 365 | Observability for Third-Party Agents

Microsoft is expanding Agent 365 to monitor AI agents that do not originate from its own ecosystem. Through the Connected Platforms, agents from Amazon Bedrock, Google Vertex AI, and Anthropic Claude Managed Agents can now be monitored in the same place as Microsoft’s own agents.

Public Preview starts at the end of October 2026, with general availability following at the end of November. It has been published as MC1480873.

For environments running AI agents across multiple platforms, this closes a visibility gap. Each synchronized agent receives an Activity tab that displays runs, sessions, tool calls, and exceptions, translated into the observability model of Agent 360.

Agent 365 reads this activity via the existing connection data; it does not execute the third-party agents itself. One metric is deliberately omitted: users cannot be authenticated for the third-party agents because these platforms do not provide Entra identities.

A single switch is crucial for governance. When creating or editing a supported connection, the setting for collecting observability data appears. For new connections, it is enabled by default, but you can disable it during setup.

For existing connections, collection remains disabled until an administrator enables it. The agents themselves continue to synchronize as before, regardless of this setting.

The data resides on the respective source platform, and that's also where the effort is incurred. Amazon Bedrock Classic requires trace collection to S3 and read access to the activity bucket. Amazon Bedrock AgentCore demands CloudWatch Transaction Search and access to the log groups. Google Vertex AI needs roles for BigQuery, Monitoring, and Cloud Trace, along with enabled telemetry.

Anthropic Claude Managed Agents require no additional setup. However, you should factor in that third-party providers may charge separate fees for tracing, storage, or queries.

Who stands to benefit from taking a closer look? Primarily organizations that operate agents not only in Copilot Studio but also concurrently in Bedrock or Vertex—and so far lacked a unified log across all platforms. From a governance perspective, the central activity view is the real advantage, as it allows tracking runs and exceptions from third-party agents without having to jump into each platform individually.

Anyone already working with multiple agent platforms today checks after launch which connections should activate the collection and aligns data collection with data protection requirements.