Microsoft’s Agent Framework 1.21 adds full-buffer Purview policy evaluation for streamed responses, alongside new agent guidance and connectors. For developers, it is a practical set of changes to how agents handle tools, data and streamed output, rather than a new model announcement.
Microsoft AI Watch analysis
What happened
The Python 1.21.0 release, published on 8 October, adds Purview policy evaluation for streamed responses and guidance for tool results. It also exposes rewritten variable-expansion arguments, adds a maximum reasoning-effort option to the OpenAI connector, and introduces an alpha Oracle native vector-store connector. Read the Python 1.21.0 release notes.
The notes also list fixes across the framework, including improved validation and media detection, more careful handling of shared file edits, and changes to how provider streams close when users stop reading them.
Why it matters
The release gives teams building agents more ways to apply policy to streamed output and connect agents to different data stores. Streamed responses arrive in pieces, so evaluating the full buffer for Purview policies is a concrete option for teams whose checks need the complete response before it is assessed.
The Oracle connector is marked alpha, so it is an early integration rather than a finished promise. The rest of the release is a varied framework update, not evidence that every application built on it will behave differently.
Our read
This is useful plumbing with a couple of changes that reach beyond tidying the toolbox. Policy evaluation for streamed responses and an expanded choice of connectors give developers specific new options to test; the alpha label is a reminder to try the Oracle integration before trusting it with anything important. A changelog can’t do the testing for you, but at least this one names the parts worth looking at.
What to watch
- How the Purview evaluation behaves in applications that stream responses to users.
- Whether the Oracle connector moves beyond alpha and gains broader support.
- How developers use the new tool-result guidance and reasoning-effort option.
Discussion spark: For agent developers, which matters more in this release: policy checks that evaluate streamed output, or new connectors that expand where agents can get data?
Sources and evidence
- python-1.21.0 (8 October 2026, 10:28 UTC)
not affiliated with or endorsed by Microsoft