Microsoft Agent Framework 1.0 shipped on April 3, 2026 — the GA merger of AutoGen and Semantic Kernel. It supports checkpointing, pause/resume, human-in-the-loop approvals, and all major orchestration patterns (sequential, concurrent, handoff, group chat, Magentic-One) for both .NET and Python. The GA removes the main reason to hold back from Trial.
Why It Moved to Trial
The April 3 1.0 release delivered:
- Stable, versioned APIs — long-term support commitment, no further breaking changes in the 1.x line
- Checkpointing + pause/resume — long-running workflows survive failures and disconnections
- Agent Harness core (released July 23, 2026) — planning, per-call history persistence, compaction, file memory, approvals, web search, and OpenTelemetry in one configurable runtime
- Native A2A and MCP support — plugs into the cross-vendor protocol stack
The prior blocker — RC status with slipping GA timeline — is resolved. The remaining limitation is that some advanced orchestration packages stay in preview even with the 1.0 core GA. The July harness release also leaves background agents, file access, automatic looping, and shell tooling opt-in with prerelease warnings, and the community ecosystem is still thinner than LangGraph/CrewAI.
August 2026: Durable Approvals and Enforcement Hooks
Python 1.14.0 adds experimental AGENT-HOOKS enforcement middleware, AG-UI workflow checkpoint/resume, Foundry state stores for sessions, checkpoints, and function approvals, plus explicit background-session release. It also preserves approval replay and conversation correlation across AG-UI runs, rejects Windows junctions during skill discovery, scopes under-specified session approvals to the current tool, and bounds compaction summaries before provider calls. Durable Task and Azure Functions integrations moved into a separately published extension repository, while Foundry hosting adopted the Responses 2.x storage model with beta breaking changes. (Python 1.14.0)
This strengthens the Trial thesis rather than promoting it: persisted approvals and checkpoint semantics are production-relevant, but hooks remain opt-in experimental and parts of the hosted storage/package boundary are still moving.
Buy vs Build
This is a build framework with a strong buy story if your organisation is on Azure. The managed hosting, enterprise SLAs, and Azure integrations are the "buy" angle — you're buying reliability and compliance, not avoiding code.
AutoGen brought: event-driven multi-agent orchestration, async execution, human-in-the-loop patterns, production use at Novo Nordisk and other enterprises
Semantic Kernel brought: production-grade LLM SDK, C#/Python/Java support, deep Azure OpenAI integration, plugin architecture
Combined, you get:
- Multi-language support (C#, Python, Java) — uniquely valuable for heterogeneous enterprise stacks
- Native Azure integration (Azure OpenAI, Azure AI Search, Azure Functions)
- Production SLAs with Microsoft support
- Compliance features important in regulated industries
- A batteries-included Harness Agent for Python and .NET, while preserving the lower-level chat-client agent for teams that want to assemble their own runtime
When to Choose It
- Your engineering team writes C# or .NET
- You're committed to Azure infrastructure
- You need Microsoft support SLAs
- You're in a regulated industry where Microsoft's compliance certifications matter (HIPAA, SOC 2, ISO 27001)
- You want checkpointing and pause/resume without building it yourself (cf. LangGraph, which has this too)
When to Choose Something Else
- Python-only stack: LangGraph or CrewAI have more ecosystem momentum and community Q&A coverage
- Not on Azure: The Azure integration is the key value prop; outside Azure, you lose most of the "buy" benefits
- Need model-agnostic + Claude: MAF works with any model, but its deep integrations favor Azure OpenAI
Key Characteristics
| Property | Value |
|---|---|
| Version | 1.0 GA (April 3, 2026); Harness core released July 23, 2026 |
| License | MIT |
| Language | C#, Python, Java |
| Provider | Microsoft |
| Checkpointing | Yes — built-in pause/resume |
| A2A / MCP | Native support |
| GitHub | microsoft/agent-framework |
| Website | aka.ms/AgentFramework/Docs |
Sources
- Microsoft Agent Framework Version 1.0 — Foundry blog — April 3, 2026 GA announcement
- Microsoft Ships Production-Ready Agent Framework 1.0 — Visual Studio Magazine — April 6, 2026
- Build Multi-Agent AI Apps on Azure App Service with MAF 1.0 — Microsoft Community Hub
- The Microsoft Agent Framework Harness is now released — July 23, 2026; default harness components and remaining opt-in preview features