Microsoft Agent Framework changelog: what's new each month
Every stable Microsoft Agent Framework release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 10 months covered; the current month updates daily.
November 2026 · so far
3 releases: 1.28.0 → 1.26.0
November brought significant security and stability improvements across the Agent Framework, with stricter defaults for sensitive operations and new capabilities for computer tool integration and structured outputs. Multiple breaking changes require configuration updates for secure hosting scenarios.
- A2A: Individual identity-parameter overloads replaced with optional A2AAgentOptions for agent construction.
- A2A hosting: Built-in server registration limited to message-based responses only.
- GitHub Copilot: File hooks disabled by default for new and resumed sessions.
- CodeAct: Host-tool exception details redacted by default in Hyperlight and LocalCodeAct.
Highlights
- Computer tool now available in Foundry with batched action support for cloud hosting.
- Structured outputs via Pydantic models and JSON Schema now supported in GitHub Copilot.
- Security hardened: exception details redacted by default, file hooks disabled, cross-workflow state isolation enforced.
- MCP discovery now bounded with explicit approval policies and authorization scope enforcement.
- Execution isolation improved with finite budgets and cancellation handling in code execution.
New
- GA computer tool through FoundryAITool.CreateComputerTool() with batched action support.
- TokenScope option in FoundryToolboxOptions for cloud-specific authentication.
- Pydantic model and JSON Schema support for structured responses in GitHub Copilot.
- Optional message_labeler callback for application-owned context provider classification.
- Task-scoped authority middleware guidance and executable example.
- Experimental AgentFactory created on-demand in declarative workflows.
Improved
- loop_return_final_only option added to create_harness_agent() for non-streaming callers.
- MCP tool and prompt discovery now bounded at 100 pages and 1,000 items per catalog.
- Framework-owned MCP request timeouts defaulted to 30 seconds.
- Hyperlight and LocalCodeAct preserve successful results while redacting exception details.
- Execution budget enforcement with 30-second default limit in Hyperlight.
- Cancelled execution properly interrupted using Hyperlight SDK interruption API.
- GitHub Copilot CLI version validation at startup.
- A2A, A2A.AspNetCore, and A2A.Grpc updated to 1.0.0-preview4 with new request-context API.
Fixed
- Live workflow agent executors and sessions now bound to single authorization scope with cross-scope rejection.
- MCP state no longer shared across workflows; fresh protocol session created per handler invocation.
- Diagnostic policy preserved in serialized execution state across Hyperlight restarts.
- Graph and functional workflow invocation spans aligned with OpenTelemetry GenAI conventions.
October 2026
10 releases: 1.23.0 → 1.24.0
October brought significant updates across both .NET and Python implementations, with major architectural improvements to Foundry hosting, new vector-store connectors, and enhanced security hardening throughout the framework.
- Better support for tool changes between runs requires code updates.
- Foundry hosting redesigned around request-scoped agent factories and persistent sandbox-isolated sessions.
- AG-UI now enforces fail-closed request trust boundaries and binds workflow continuations to resolved thread ID.
- AG-UI prevents client Shared State and stored Thread Snapshots from replacing tool-approval state.
Highlights
- Redesigned Foundry hosting around request-scoped agent factories and persistent sandbox-isolated sessions for better scalability and isolation.
- Added native computer-use support to Responses clients and Foundry project embeddings.
- Introduced new DuckDB and SQL Server native vector-store connectors alongside the TypeSafe AI connector.
- Enhanced approval and MCP runtime context handling with stricter validation and fail-closed defaults.
- Added MCP lifecycle management improvements and new authorization requirements for AG-UI endpoint registration.
New
- Native workflows can now be hosted over Foundry Responses and Foundry Invocations.
- DuckDB and SQL Server native vector-store connectors and TypeSafe AI connector.
- Response-stream gates and buffering for better control over agent responses.
- Session-scoped file-access isolation for improved security.
- Opt-in synthetic trailing user turn for providers that return empty text after assistant-ending prompts.
- MongoDB agent RAG sample demonstrating knowledge retrieval patterns.
- Alpha Oracle native vector-store connector.
- Alpha NVIDIA NIM embedding client with OpenAI-compatible request-options extension point.
Improved
- Hardened approval and MCP runtime context handling with stricter validation and fail-closed behavior.
- Tool-approval requests that never receive a response are now automatically rejected.
- Preservation of explicit null session state across roundtrips.
- Better support for tool changes between runs.
- Citations and error details now included in Purview policy checks.
- Experimental agent-triggered in-run compaction and skill unloading strategies.
- Async conditions now supported in switch-case workflow edge groups.
- Configurable service URLs added to Ollama and OpenAI chat clients.
Fixed
- Fixed declarative workflow HTTP variables handling.
- Fixed workflow topology edge multiplicity comparison.
- Fixed reasoning remnants and handoff bounce-back issues.
- Corrected endpoint override instructions in Responses samples.
- Fixed Anthropic client behavior to honor allow_multiple_tool_calls when tool_choice is unset.
- Fixed Copilot session hooks composition.
September 2026
5 releases: 1.17.0 → 1.22.0
September brought significant expansions to vector storage capabilities and workflow management, along with improved OpenAI SDK compatibility and new observability features. The month included several breaking changes to agent middleware and file access APIs.
- Agent middleware inputs reverted to sequence-only; agent-hooks core extra removed.
- AgentFileStore line-numbering contract changed; fileaccessreadlines added.
- Azure.AI.OpenAI dependency removed (.NET).
Highlights
- New vector-store abstractions with implementations for Azure AI Search, Redis, MongoDB, Azure DocumentDB, and Azure Cosmos DB.
- Added maximum-duration bounds and stop-reason signals to tool invocation loops for better control.
- Support for OpenAI SDK 3.x while maintaining backward compatibility.
- MCP Host-history conversion for persisted AG-UI conversations.
- Aspire traces now displayed in the developer UI for enhanced observability.
New
- Shared vector-store abstractions with portable filters and in-memory implementation.
- Azure AI Search vector-store connector.
- Redis HASH and JSON vector stores.
- MongoDB alpha vector-store connector.
- Azure DocumentDB alpha vector-store connector.
- Azure Cosmos DB NoSQL vector-store implementation.
- Stable names for built-in orchestration workflows with checkpoint type registration.
- Sequential function call invocation option.
Improved
- Mistral SDK migrated to official SDK for chat and embedding clients.
- Agent middleware inputs restored to sequence-only mode with better performance.
- Provider refusals now preserved as marked text across conversation history.
- Function call approvals bound to stable occurrences while preserving legacy compatibility.
- Mixed workflow invocation keyword arguments now supported.
- CodeAct tool parameter schemas exposed with configurable compact or JSON descriptions.
- Developer UI now supports emitmessagessnapshot configuration for suppressing terminal message snapshots.
- Foundry hosting improved with explicit Agent Server or agent-owned model history selection.
Fixed
- Race condition in workflow formula state.
- Duplicated conversation replay issue in Foundry-hosted agents.
- Delimiter value inspection before adding headers.
- Multiple typos in architecture decision records.
August 2026
7 releases: 1.17.0 → 1.20.0
August brought significant enhancements across the Microsoft Agent Framework, including new AI provider integrations, workflow improvements, and foundational changes to support more resilient and flexible agent deployments. The month saw multiple releases focused on expanding platform capabilities and hardening production readiness.
- Renamed AgentIsolationKeyProvider (version 1.18.0).
- Consolidated OpenTelemetry GenAI semantic-convention support with explicit message-event controls, requiring updates to observability configuration (version 1.15.0).
Highlights
- Added native Mistral AI chat client support with streaming, tools, and structured output capabilities.
- Introduced workflow checkpoint creation, resumption, and steering support for more resilient agent operations.
- Expanded foundry-based hosting with provider-scoped state stores and session continuity features.
- Added optional agent-generated UI support and improved streaming tool-call handling for better user interfaces.
- Introduced configurable timeouts and enhanced OpenTelemetry configuration for better observability and control.
New
- Mistral chat client with native support for chat, streaming, tools, structured output, and embeddings.
- Experimental agent-hooks enforcement middleware for extensible interception contracts.
- Workflow checkpoint creation and resume capabilities in Agent Framework UI.
- BackgroundAgentsProvider.releasesession() method for safely cancelling work and releasing session runtime state.
- Provider-based Foundry state stores for managing agent sessions, checkpoints, and function approvals.
- Optional A2UI support for agent-generated interfaces with preserved streaming tool-call indices.
- MiddlewareFailure as a first-class fatal signal for function middleware.
- Mem0Sharp integration for in-memory storage in agent samples.
Improved
- Declarative workflows now fail when an agent returns an error instead of continuing.
- Tool-approval auto-approval loop is now bounded to prevent runaway approvals.
- File skill discovery is hardened for better reliability.
- Hosted agents now have a single source of conversation history.
- Usage metrics are aggregated across looping agents and chat clients.
- Foundry Hosted Agents now include steering, retry, and recovery support with long-running workflow samples.
- Run continuity improved by avoiding emission of unchanged predictive state snapshots.
- OpenTelemetry GenAI semantic-convention support consolidated around stable and experimental modes.
Fixed
- Fixed handoff orchestration sample not responding to user input in .NET.
- Fixed snakecase argument names in Harness file tool descriptions.
- Fixed workflow checkpoints being mutatable outside storage implementations.
- Preserved logprobs field in Responses API.
- Preserved unmapped provider finish reasons in Anthropic and Mistral integrations.
- Honored cancellation for Foundry-hosted workflow responses.
July 2026
10 releases: 1.12.0 → 1.13.0
July brought significant stability improvements and graduation of experimental features to production readiness. The framework expanded capabilities for message injection, skill source management, and multi-model support across .NET and Python platforms.
- Foundry experimental flags aligned to MAAI001 for MAF-specific APIs.
- Caching extracted from AgentSkillsProvider into separate CachingAgentSkillsSource class.
- File editing tools added and FileAccess/FileMemory store API changed.
- Message injection graduated from experimental, changing API surface.
Highlights
- Message injection and todo/agent mode providers graduated from experimental status to stable.
- Added progressive MCP disclosure for on-demand tool schema discovery while maintaining security boundaries.
- New Azure Cosmos DB semantic-memory provider with fact extraction and user profiles.
- GitHub Copilot agent and AG UI packages promoted to stable release.
- Enhanced observability with OpenAI cache-write token usage details and prompt cache support for GPT-5.6 models.
New
- Message injection middleware allows tools and host code to enqueue messages into active runs within the same session.
- Progressive MCP disclosure enables agents to discover, load, and unload tool schemas on demand.
- Azure Cosmos DB semantic-memory context provider with fact extraction and user profile support.
- File editing tools and file access/memory store API for enhanced file operations.
- App-owned A2A and MCP hosting helpers for exposing agents and workflows as native MCP tools.
- Skill approval options and ability to disable approval for specific tools like FileAccessProvider.
- Custom argument parsing for inline skill scripts to handle non-conforming tool-call argument shapes.
- Bounded, in-memory archive skill discovery for MCP sources with archive controls in FoundryToolbox.
Improved
- Skill source caching extracted into CachingAgentSkillsSource with configurable TTL for cache expiration.
- Foundry Hosting now supports per-user session isolation and Responses v2 protocol fast-fail.
- SkillsSourceContext threaded through skills source pipeline for context-aware filtering and agent attribution.
- MCP session provider option added to MCPSkillsSource for resolving sessions on every fetch while preserving cached skills.
- HITL response-URL addressing for requests raised from inside workflows.
- Workflow credential selection hardened for improved security.
- Support for ephemeral per-request instructions in the Responses API.
- Asynchronous credential support in FoundryToolbox.
Fixed
- CosmosChatHistoryProvider now correctly omits TTL when MessageTtlSeconds is null.
- Fixed CompactionMessageIndex.IsSummaryMessage logic.
- Workflow session bug resolved.
- InMemoryChatHistoryProvider no longer persists when service stores history.
- Stateless replay of reasoning-paired tool calls fixed across multiple frameworks.
- Gemini 3 thought signatures now preserved across function-call replays.
June 2026
11 releases: 1.6.2 → 1.10.0
June brought major expansions to skills and tool management across both .NET and Python, with MCP-based skills discovery becoming a core capability and significant improvements to workflow declarative support. Tool approval and sampling guardrails were added to enhance safety.
- AgentSkill API refactored to async resource and script lookup.
- Code-Gen support removed from Declarative Workflows.
- File-access tools now require approval by default with read-only auto-approval in Python.
- All AgentSkillsProvider tools now require approval by default in .NET.
Highlights
- MCP-based skills discovery now available in both .NET and Python for flexible tool integration.
- Tool approval middleware and guardrails added to control agent tool execution and sampling behavior.
- Declarative workflow packages promoted to stable release with improved checkpoint support.
- Long-running task support for MCP client tools enabling more complex agent workflows.
- AgentSkill API refactored to async resource and script lookup for better performance.
New
- MCP-based skills support (skill-md type) for Foundry Toolbox tools.
- Tool approval middleware with integration into harness agent and approval callbacks.
- AgentLoopMiddleware for re-running agents in a loop.
- AgentFileStore and FileAccessProvider for file access operations.
- A2AAgentSession with reference task IDs and input-required support.
- Foundry Adaptive Evals integration for rubric generation.
- AG-UI thread snapshot persistence and hydration (opt-in).
- Background agent support to harness agent with automatic provider resolution.
Improved
- HarnessConsole rendering performance improved with reduced flickering.
- Parallel tool call rendering fixed in AGUI translation layer.
- MCP tool kwargs filtered to declared parameters via allowlist for safer execution.
- OpenAI provider response compatibility improved for non-OpenAI providers.
- Code interpreter history chunks coalesced for efficiency.
- Sync tools now run off the event loop.
- MCP servers auto-labelled based on hints and GitHub interface labels.
- Agent telemetry enhanced to capture context provider instructions.
Fixed
- MessageId now properly populated from TaskStatusUpdateEvent in A2AAgent.
- Magentic agent replies now correctly shared across team.
- Checkpoint resumption issue after package version upgrade resolved.
- Per-service-call history persistence fixed with server-storing clients.
- AGUI hosting and workflow bugs resolved.
- Duplicate render and text input clear bugs fixed with improved guardrail error messaging.
May 2026
8 releases: 1.4.0 → 1.7.0
May brought major feature additions across the framework, including new skill definition capabilities, improved workflow support, and experimental hosting features. The month focused on expanding agent functionality while strengthening core infrastructure and compatibility.
- Instrumentation is now enabled by default across core and Foundry packages.
- Experimental skills API: File skill folder discovery aligned with agentskills.io specification.
- Experimental skills API: Skill spec metadata extracted into SkillFrontmatter.
- String array arguments now supported for file-based skill scripts.
Highlights
- New ClassSkill and file-based skill definitions with automatic method discovery and declarative metadata support.
- Shell tool added with support for both local and Docker execution.
- Experimental hosted agents framework with RAG, skills, and memory samples.
- Improved workflow handling including HttpRequestAction support and better intermediate output processing.
- New HarnessAgent and background-agents harness provider for testing and development.
New
- ClassSkill for class-based skill definitions with declarative metadata and automatic method discovery.
- Shell tool supporting local and Docker execution.
- Experimental session-mode, todo-list, and memory harness context providers.
- HttpRequestAction support in declarative workflows.
- Message filtering to exclude non-portable content types.
- IChatMessageInjector for message injection during function loops.
- HarnessAgent and background-agents harness provider.
- Experimental hosted tool factories on FoundryChatClient.
Improved
- Support for GPT-5 verbosity option and OpenAI/Gemini tool choice configuration.
- Agents now notified of external AgentModeProvider mode changes.
- Todo provider multithreading improvements with injection into message list.
- YAML block scalar parsing support in file skills and frontmatter.
- Better handling of intermediate outputs across workflows and orchestrations.
- Foundry Toolbox MCP now shows more authentication methods.
- Tool result display channel added to ag-ui.
- Improved DevUI with various enhancements.
Fixed
- JSON serialization issue with MultiPartyConversation.
- QuestionExecutor looping after GotoAction re-entry in declarative workflows.
- Missing Workflows shared sources in solution.
- HostedAgents strict URL routing in FoundryAgent.
- Per-service input persistence on stream error.
- MSBuild parallel-output races in Foundry.Hosting tests.
April 2026
12 releases: 1.0.0 → 1.2.2
April brought Microsoft Agent Framework to production stability with the 1.0.0 release, followed by rapid feature additions and improvements across multiple versions. The month saw significant expansion in platform support, new integrations, and hosting capabilities.
- Dependency floors now require released >=1.0.0,<2 packages, breaking compatibility with older RC installs.
- Removed deprecated BaseContextProvider and BaseHistoryProvider aliases.
- Removed deprecated text parameter from Message constructor.
- Removed OpenAIAssistantClientExtensions class (.NET).
Highlights
- Production release of Agent Framework 1.0.0 with stable APIs and dependency requirements.
- Added support for Google Gemini and Azure Content Understanding for automated file analysis.
- Introduced Foundry Hosted Agents support enabling agent hosting and sessions.
- Launched functional workflow API and improved workflow context management.
- Added A2A bridge support and Azure Cosmos DB checkpoint storage for Python workflows.
New
- GeminiChatClient for Google Gemini integration.
- Hyperlight CodeAct package and documentation.
- Experimental file history provider.
- Functional workflow API for Python.
- Azure Content Understanding context provider for auto-analyzing file attachments.
- OpenTelemetry integration for GitHubCopilotAgent.
- Hosted Durable Workflow support with full conversation history propagation.
- Streaming support for A2A agent handler.
Improved
- File skills terminology standardized to use 'directory'.
- Handoff workflow context management improved.
- Resilience of verify-samples improved with separate building and better evaluation.
- AG-UI demo improvements and reasoning role parsing.
- OpenAI Responses streaming to propagate created-at from final response event.
- Observability spans now correctly nested when using streaming.
- Checkpoint deserialization restricted by default for security.
- Removed [TOOLBOXES] warning from FoundryChatClient calls.
Fixed
- Off-thread RunStatus race condition where GetStatusAsync could return Running after ResumeAsync halts.
- OpenAI embedding client now uses AsyncOpenAI for /openai/v1 endpoints.
- File search citations breaking assistant-message history roundtrip.
- Duplicate CallIds causing Handoff Message Filtering to fail.
- DevUI frontend streaming memory growth.
- Telemetry sample bug.
February 2026
2 releases: 1.0.0b260210 → 1.0.0b260212
February brought Microsoft Agent Framework to version 1.0.0 beta with major foundational changes. The month focused on streamlining the API, adding long-running agent support, and restructuring core types for simplicity.
- ChatAgent renamed to Agent, RawChatAgent to RawAgent, ChatMessage to Message, ChatClientProtocol to SupportsChatGetResponse.
- Single getresponse and run API replaces previous separate methods.
- Context provider pipeline wiring removes old types; all consumers updated.
- Checkpoint format and encode/decode mechanism changed.
Highlights
- Long-running agents and background responses now supported with continuation tokens.
- Core types renamed for a simpler API: ChatAgent becomes Agent, ChatMessage becomes Message, and others.
- Unified getresponse and run API replaces previous separate methods.
- Streaming support added for code interpreter deltas.
- Tool handling centralized and simplified with new @tool decorator schema support.
New
- Long-running agents with background responses and ContinuationToken support.
- Streaming support for code interpreter deltas.
- Explicit input, output, and workflowoutput parameters for @handler, @executor, and requestinfo decorators.
- Explicit schema handling in @tool decorator.
- New session and context provider types.
- AzureOpenAIResponsesClient can now be created with Foundry project endpoint.
- Centralized tool result parsing through FunctionTool.invoke().
Improved
- Context provider pipeline now wired throughout the framework.
- Checkpoint system refactored for better encode/decode handling and format.
- HostedTool classes replaced with tool methods for simpler implementation.
- Settings management moved from Pydantic Settings to TypedDict with loadsettings().
- Python samples restructured into progressive 01-05 layout for better learning.
- Orchestration examples reorganized with AzureOpenAIResponsesClient adoption.
Fixed
- Workflow and orchestration bugs fixed in samples.
- Purview package now has comprehensive test coverage.
January 2026
2 releases: 1.0.0b260128 → 1.0.0b260130
January brought significant structural improvements to the Agent Framework, including breaking changes to core APIs and expanded test coverage across packages. The month focused on stabilizing the foundation with new type safety features and Claude SDK support.
- @aifunction decorator renamed to @tool.
- AIFunction renamed to FunctionTool.
- Github renamed to GitHub in agent-framework-github-copilot.
- GroupChatBuilder and MagenticBuilder now require factory pattern usage.
Highlights
- The @aifunction decorator is now @tool, and AIFunction is now FunctionTool.
- GroupChatBuilder and MagenticBuilder now use the factory pattern.
- BaseAgent implementation added for Claude Agent SDK.
- Generic types added to ChatOptions and ChatResponse/AgentResponse for better Response Format handling.
- Test coverage expanded significantly across core, utilities, observability, and Azure AI packages.
New
- BaseAgent implementation for Claude Agent SDK.
- Core types and agents unit tests.
- Core utilities unit tests.
- Observability unit tests.
- Generic types in ChatOptions and ChatResponse/AgentResponse for Response Format.
Improved
- GroupChatBuilder and MagenticBuilder now use factory pattern.
- AzureAI package test coverage improved.
- Durabletask package updated.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Microsoft Agent Framework: pricing, features and alternatives · All changelogs