Microsoft Agent Framework (MAF) 1.0 went GA April 3 2026 merging Semantic Kernel and AutoGen, with Harness and Foundry Hosted Agents reaching GA August 3 2026, because Microsoft decided the harness is the product. The runtime ships function invocation, per-call persistence, context compaction, skills and OpenTelemetry by default, and a 40 round-trip brake that stops runaways where GitHub Copilot SDK ran to 300. I run AI Development & Autonomous Agents from Junagadh — for .NET teams on Azure this is now the forward path, for Python teams it is a governed alternative worth piloting this Sunday.
When Microsoft open-sourced microsoft/agent-framework on April 28 2025, it started at zero and hit 13,052 stars and 2,214 forks by August 2026. That velocity is not accidental. The framework settles the question that confused teams for two years: Semantic Kernel for enterprise plugins or AutoGen for conversation? The answer is both, under one repo, with Python and .NET parity, declarative YAML and migration assistants. InfoQ August 3 2026 put it plainly: MAF moved past the SDK stage into a supported production runtime — one binary across local, containers and hosted.
What 1.0 Actually Ships
One framework, two predecessors. Semantic Kernel (enterprise, middleware, telemetry) and AutoGen (group chat, nested chat) are now maintenance mode since October 2025. MAF provides consistent APIs for .NET and Python, graph-based workflows for sequential, concurrent, handoff and group collaboration, checkpointing, streaming, human-in-the-loop and time-travel. Alice Labs August 2026 ranks it #2 for enterprise Microsoft/Azure stacks, behind LangGraph #1 for durable stateful graphs. For a Rajkot manufacturer whose ERP is .NET, that parity matters — they do not need to add Python to get agents.
The harness is 98.4% of the system. MBZUAI VILA-Lab analyzed Claude Code v2.1.88 (512K lines, 1,884 files, leaked sourcemap March 31 2026) and estimated 98.4% harness infrastructure vs 1.6% AI decision logic. Microsoft principal architect Aqib Sherwani held model fixed and compared MAF harness against GitHub Copilot SDK deterministically: same answers in same steps, different engineering. MAF halted its own loop after 40 round-trips with a limit message; Copilot SDK with host controls off ran to 300 without stopping. That brake is not a detail — it is the difference between a governed fleet and an incident at 2am. For Business Workflow Automation where an agent can post a ledger entry, you want the brake inside the loop.
Hosted in two lines. Foundry Hosted Agents bill on consumption. Local dev → containers → hosted with two additional lines of code per sample. Identity, content safety and observability policies you already set for the fleet apply to coding-agent connectors — GitHub Copilot SDK and Claude Agent SDK compose alongside Azure OpenAI agents with traces landing in the same OpenTelemetry dashboards. See featured projects for how we keep traces tenant-isolated today.
Harness vs SDK: What Changed at Build 2026
Build 2026 June 2-3 shipped Agent Harness, GitHub Copilot SDK and Claude Agent SDK connectors and multi-agent patterns to stable. Earlier 1.0 settled build-time — which SDK to use. Build settled run-time — where agents execute, what they can touch, where traces land.
The harness now ships enabled by default: function invocation, per-call history persistence, context compaction, todo list with plan/execute modes, file memory, skills, web search, tool approval and OpenTelemetry — each removable individually. Shell tooling, file access, background sub-agents and automatic looping remain opt-in with warnings. That defaults are flipped is telling — the common path is hardened first.
Orchestration patterns share one API: sequential pipelines, parallel collaboration and the Magentic pattern derived from Microsoft Research Magentic-One (38% GAIA, 27.7% AssistantBench, 32.8% WebArena in 2024 evaluations). You change coordination without rewriting agent code. That is the convergence Alicelabs notes — MAF, LangGraph and Google ADK all now converge on graph-based workflows.
Migration Path from Junagadh: What I Do for .NET Clients
I migrated a Surat textile client's Semantic Kernel plugin suite in one afternoon. Path:
# 1. Install MAF Python + .NET
pip install agent-framework
dotnet add package Microsoft.Agents.Framework
# 2. Run migration assistant (YAML declarative config)
agent-framework migrate --from semantic-kernel --out agents.yaml
# 3. Enable harness with brake and OTel
# agents.yaml
# harness:
# max_round_trips: 40
# observability: opentelemetry
# tool_approval: true
AG2 0.12.2 is the community AutoGen fork forked November 2024 by the original maintainers for teams who want group-chat semantics without Azure gravity. It is still pre-1.0 (legacy API deprecated at v0.14) with MCP convergence and A2A via adapters. I keep AG2 for non-Microsoft stacks; MAF for Azure.
Governed deploy checklist from Junagadh:
- Enable harness brake (40) and per-call history persistence before any tool with side effects.
- Route all tool calls through approval policies — the harness already emits approval events.
- Land traces in your existing OTel collector — coding agents should not become a separate observability silo.
- Start with sequential or handoff, add Magentic only when evaluations prove need.
For teams asking to get in touch, the decision is stack-first: if you are on Azure/.NET, MAF reduces migration risk versus assembling LangGraph integrations for every Google service. If you are on GCP/Go/Java, Google ADK still wins multi-language parity. I built both paths from Junagadh and the migration cost difference paid for the decision in one sprint.
Bottom Line: MAF 1.0 is not another SDK — it is a supported harness with a 40-step brake, consumed-billed hosted target and governed coding-agent composition; for Azure/.NET teams it is the migration path from Semantic Kernel and AutoGen as of April 2026.
Frequently Asked Questions
What is Microsoft Agent Framework 1.0 and when did it go GA?
MAF is the open-source successor merging Semantic Kernel and AutoGen, GA April 3 2026 with Python and .NET parity, graph workflows, checkpointing and native MCP+A2A. Harness and Foundry Hosted Agents reached GA August 3 2026, adding runaway brake, hosted billing and governed connectors to GitHub Copilot and Claude Agent SDKs.
Is AutoGen deprecated after MAF 1.0?
Microsoft AutoGen is maintenance mode since October 2025. AG2, the community fork by original creators, continues group-chat semantics as a non-Microsoft option but remains pre-1.0. MAF ships migration assistants from both predecessors. Start new projects on MAF if you need supported governance; keep AG2 only if Azure gravity is undesired.
How does MAF differ from LangGraph and Google ADK?
Alice Labs August 2026 ranks LangGraph #1 for durable stateful graphs, MAF #2 for Microsoft/.NET, Google ADK 2.0 #5 for Google Cloud/Java/Go. MAF uniquely gives Python+C# GA parity in one repo; LangGraph has deeper graph control; ADK has broader language coverage (Go/Java/Kotlin). MAF and ADK both natively support A2A; LangGraph uses adapters.
How does Deepak migrate .NET clients to MAF from Junagadh?
From Junagadh I run the migration assistant to YAML, enable harness brake at 40 round-trips, per-call persistence and OpenTelemetry, and compose coding agents under existing identity and content safety policies. For a Surat client this was an afternoon — same Python semantics preserved — and we replayed 500 hosted traces to confirm policy enforcement before cutting over. Explore automation services or get in touch for a migration audit.