Microsoft Agent Framework — python-1.20.0 release
Release notes for Microsoft's Agent Framework (Python and .NET SDKs), covering versions python-1.16 through python-1.20 and dotnet-1.19 through dotnet-1.23. Changes span native vector-store connectors (MongoDB, Qdrant, PostgreSQL/pgvector, DuckDB, SQL Server, Azure Cosmos DB, Azure AI Search), Foundry hosting redesigns around request-scoped agent factories and durable Invocations, MCP skill/session security hardening (approval binding, security labels, ZIP-only skill archives), Magentic multi-agent orchestration and handoff routing, and checkpoint/replay correctness fixes across workflows.
What it claims
This is five Python releases (1.16.0 on 2026-08-27 through 1.20.0 on 2026-10-02, roughly weekly) and five .NET releases (1.19.0–1.23.0). The summary lists the themes. What it leaves out is the upgrade cost: about a dozen changes are marked [BREAKING], several defaults now fail closed, Python-version and dependency floors have moved, and some approval-bypass and data-leak bugs were fixed. Anyone pinned below python-1.18 or dotnet-1.21 should read the breaking list below before bumping.
How to use it
Python breaking changes, by release
- 1.17
- Agent middleware inputs are sequence-only again.
- The experimental
agent-hookscore extra is removed. - 1.18
- Lab is removed from
core[all]and has isolated dependency resolution. SecretStringis now a masked wrapper, not astrsubclass. This affects credential handling in Anthropic, Azure AI Search, Cosmos, Bedrock, Copilot Studio, Foundry, Gemini, Mem0 and OpenAI.- GitHub Copilot workspace file hooks are now opt-in via
enable_file_hooks. - foundry-hosting (beta) restricts checkpoint deserialization by default. Applications must register any extra checkpoint types.
- Ranged file reads move the line-numbering contract onto
AgentFileStore(experimental). - File, memory, session and skill path normalization is centralized.
- 1.19
- HTTP cookie persistence must now be explicit (a2a, ag-ui, core, declarative).
- MCP skill archives are ZIP-only, and the MCP sampling callback is deprecated.
- Declarative MCP sessions are scoped per invocation.
- Redis history keys are scoped by provider and session identity (beta). Existing stored history keys will differ.
- 1.20
- Foundry hosting is redesigned (request-scoped factories, sandbox-isolated sessions, durable Invocations).
- Declarative workflow state lookup and PowerFx bounding change (experimental).
- Gemini moves to the
GOOGLE_*settings namespace and adds Gemini Embedding 2.
Python behavior changes not marked breaking
- Approvals now fail closed when bindings or agent-tool approvals can't be resolved.
- Local approvals must be session-backed.
- Matched function results are treated as terminal.
- Tool diagnostics are internal by default (1.19).
FunctionInvocationContextinjection for skill scripts is opt-in.
Python dependency and runtime support
| Item | Status |
|---|---|
| Python 3.15 | Supported via msgspec 0.22. Postgres and Hyperlight resolution restored. Redis capped at <3.15 by redisvl |
| Python 3.14 | Enabled for Hyperlight via Hyperlight Sandbox 0.6 |
| OpenAI SDK | 3.x supported (1.17) |
| Azure AI Projects | Through 2.7 |
| Azure Monitor OpenTelemetry | ≥1.8.10 required by Foundry for trace propagation |
| FastAPI | Through 0.141 (ag-ui, devui) |
| Mistral | Migrated to the official Mistral SDK |
all meta-package | Bounded to the current release cohort, to stop pip backtracking |
Vector-store maturity. The page labels the MongoDB, Azure DocumentDB, Qdrant and PostgreSQL/pgvector connectors alpha. It gives no maturity label for DuckDB or SQL Server. The summary does not mention the Redis HASH/JSON stores, the in-memory store, or the shared portable-filter abstractions; all were added in 1.18.
.NET breaking changes
- 1.23
- Azure.AI.Projects 3.0.0-beta.3, OpenAI 2.14.0, MEAI 10.10.1.
- Configuration keys now use an allow list.
- Approval response binding is enforced consistently.
- DevUI approval continuation changes.
- Tool changes between runs are handled differently.
- AGUI SDK packages are now 1.0.0.
- 1.22
AgentSessionStoreis promoted intoAgents.AI.Abstractions(preview).- Approval binding is required for replay.
- MCP sessions are scoped per invocation.
- Foundry delegated user identity is sticky on
AgentSession. - New
AsIChatClientextension exposes anAIAgentas anIChatClient. - 1.21
- The
Azure.AI.OpenAIdependency is removed. AWSSDK.Extensions.Bedrock.MEAIis replaced byAWS.Bedrock.MEAI.file_access_read_linesis added.- A2A run modes are clarified.
- The LocalCodeAct subprocess environment is isolated.
- 1.19
- MCP long-running tasks migrate to the 2026-07-28 Tasks extension.
Security-relevant fixes
These argue for upgrading rather than staying pinned:
- .NET 1.23 (#8403): the invoke-function tool could bypass approval.
- Python 1.20 (#8972): MCP security labels are now applied before dynamically discovered tools become callable.
- .NET 1.21 (#8164): headers are no longer forwarded on redirect. The OpenAI hosting store is scoped by isolation key.
- Python 1.19 / .NET 1.21–1.22:
- MCP skill archive digests are verified.
- MCP requests and sessions are scoped to the invocation's identity and origin.
- Skill paths are revalidated before use.
- Hosted session key boundaries are preserved.
Limits
- The fetched text is truncated partway through the python-1.20, 1.19 and 1.18 "Fixed" sections, so those fix lists are incomplete.
- The .NET notes are auto-generated commit and PR lists with no migration guidance.
- Some CI checks were disabled when these releases were cut:
- OpenAI integration tests were temporarily skipped because "the CI API key is invalid" (#8767, dotnet-1.23).
- A flaky steering-harness session-reuse test was skipped (#8377, dotnet-1.22).
- Several breaking changes affect packages labeled beta or experimental (foundry-hosting checkpoints, Redis history, declarative, file store). Expect further churn there.
Organisations named here
Technologies named here
- Python
- .NET
- Microsoft Agent Framework
- Gemini Embedding 2
- Model Context Protocol
- Azure AI Foundry hosting
- DevUI
- CodeAct sandbox
- AG-UI integration
- GitHub Copilot SDK integration
- Mistral SDK integration
- Microsoft Purview integration
- Redis vector stores and history provider
- MongoDB connector
- Qdrant connector
- PostgreSQL/pgvector connector
- DuckDB connector
- SQL Server connector
- Azure Cosmos DB / DocumentDB connectors
- TypeSafe AI connector
- Magentic orchestration
- ChatKit integration
- A2A integration
- Bedrock connector
- Azure AI Search connector
- MongoDB
- Redis
- PostgreSQL
- Qdrant
- DuckDB
- Gemini
- AWS
- GitHub Copilot
- OpenAI API
Techniques named here
- CodeAct sandboxing
- Magentic orchestration
- MCP skill archives and security labels
- Approval binding
- History compaction
- Workflow checkpointing
Related in agents



