Microsoft Agent Framework 1.0 Unites Semantic Kernel, AutoGen

What shipped

Microsoft has declared its Agent Framework production-ready. Version 1.0 arrived on April 7, 2026, bringing together two agent projects that had developed side by side for more than a year: Semantic Kernel, aimed at enterprise orchestration, and AutoGen, which grew out of multi-agent research 1. Microsoft describes the release as stable, with fixed APIs and long-term support for both .NET and Python 5. It follows a public preview first shown in October 2025 53.

The merger changes how the code is organized. Semantic Kernel becomes the base layer, and its kernel, plugin model and connectors carry over largely unchanged 1. AutoGen's ideas about multi-agent orchestration have been rebuilt on top of that base as a graph-based workflow engine 1. The AutoGen name remains as a compatibility layer, but new work now happens in a single repository at github.com/microsoft/agent-framework 1. Visual Studio Magazine describes the result as a successor path that keeps the underlying ideas of both projects rather than discarding them 5.

The migration is real work

This is more than a rebrand. Import paths and the API surface have changed 1. The bigger change is conceptual. AutoGen modeled multi-agent systems as conversations, with agents exchanging messages in chat threads. Agent Framework models them as directed graphs, where agents are nodes and transitions between them are explicit 1. Teams running AutoGen in production will have to rework how their systems are structured, not just rename packages.

Practitioner guidance recommends caution. One migration guide suggests starting with an inventory of existing agents rather than swapping dependencies. It recommends moving simple single-agent paths first, keeping tool contracts that already work, and then migrating state, memory, multi-agent workflows, checkpointing, observability and deployment in stages with rollback plans 2. The guide gives an example of each end of the spectrum. A support-triage bot with three tools and no persistent memory is an easy early candidate. A regulated approval agent that keeps threaded state and relies on exact streaming behavior should move more slowly 2. Its conclusion is reasonable: use Agent Framework 1.0 by default for new Microsoft agent projects, and migrate existing ones in stages 2.

Observability as the selling point

The more interesting part of the release is how agents are monitored once they are running. Microsoft's Azure guidance presents Agent Framework as Azure-native and treats agents as managed workloads that platform teams deploy, secure and monitor through the Azure AI Foundry control plane 4. Tracing is switched on automatically without code changes. It records agent reasoning, tool calls and interactions between agents in Application Insights 4. This matters because agents do not run in a straight line. They branch on intermediate reasoning, call several tools and coordinate with other agents, so ordinary request logs often fail to explain what went wrong 4.

The tracing is built on OpenTelemetry alongside Azure Monitor 4. Microsoft's own comparison places its framework next to Semantic Kernel, LangChain, LangGraph and the OpenAI Agent SDK, scoring each on how much of an agent's decision flow it exposes 4. Visual Studio Magazine also notes that deeper tooling and observability work, along with a DevUI and hosted-agent integration, is shipping as preview features alongside the stable core 5.

Open protocols, Azure gravity

Microsoft is also stressing interoperability. The framework supports OpenAPI for connecting arbitrary APIs, Agent2Agent (A2A) for collaboration across runtimes, and Model Context Protocol (MCP) for discovering tools at runtime 3. It also supports orchestration patterns such as Magentic One 3. KPMG, an early customer, says the open-source connectors let its Clara AI agents work with Azure AI Foundry as well as with outside systems and language models 3.

On balance, the openness is real, but it has limits. Adopting OpenTelemetry, MCP and A2A means that traces, tools and agent-to-agent messages use formats other platforms can read, so the framework is not a closed system. However, the strongest features are concentrated in Azure: zero-code tracing, governance, compliance and the Foundry control plane 43. The pitch is "build anywhere, operate best on Azure" 3. A team can run Agent Framework elsewhere and export standard telemetry, but it gives up the managed experience that Microsoft is using to sell the release.

Why it matters

For developers in Microsoft's ecosystem, choosing between Semantic Kernel and AutoGen is no longer necessary. There is now one supported path with long-term commitments 51. That clarity is useful, but it comes with a migration cost, particularly for teams with AutoGen code built around conversations 12. For the wider agent market, the more important question is whether OpenTelemetry-based tracing becomes a common standard across frameworks. Microsoft's decision to compare itself directly with LangChain, LangGraph and OpenAI's SDK on observability 4 suggests it expects to compete on how well agents can be monitored in production as well as on how they are built.