For anyone who spent 2025 trying to pick between Microsoft’s two agent toolkits, the decision just got simpler. In October 2025, Microsoft merged AutoGen with Semantic Kernel into a single, unified Microsoft Agent Framework, with general availability targeted for the end of Q1 2026. Two overlapping projects becoming one is rarely glamorous news, but for a team trying to standardise on a stack it is exactly the kind of boring that saves money — fewer competing abstractions to learn, and a clearer path from prototype to production.
Why two became one
AutoGen and Semantic Kernel grew up solving adjacent problems. AutoGen leaned into multi-agent conversations — agents talking to agents to get work done — while Semantic Kernel focused on wiring language models into existing applications with a more conventional, enterprise-friendly structure. The trouble with two first-party frameworks from the same company is that builders end up asking the wrong question: not “how do I solve this?” but “which of these two do I bet on?” Every hour spent on that comparison is an hour not spent shipping.
Merging them removes that fork in the road. A unified framework means:
- One set of abstractions to learn, document, and hire for — rather than two mental models that nearly, but not quite, overlap.
- A clearer migration path for anyone who already built on either tool, instead of being stranded on a deprecated track.
- Less duplicated effort in tutorials, community support, and integrations, which tends to follow consolidation.
Two overlapping first-party frameworks becoming one is the kind of boring news that quietly saves a small team weeks of indecision.
GA is the signal that matters
The phrase to watch is “general availability.” A framework in preview is a moving target — APIs shift, features come and go, and building on it is a calculated gamble. GA, targeted here for the end of Q1 2026, is Microsoft signalling that the surface is stable enough to commit to. For a middle manager weighing whether to let a team standardise on a tool, that distinction is the whole ballgame. You can justify building a real internal process on a GA framework in a way you simply can’t on a preview.
That said, even-handedness is warranted. Open-source agent tooling across the board is still maturing, and Microsoft’s framework is one strong option among several, not an automatic default. Consolidation reduces churn, but it doesn’t eliminate it — early GA releases still have rough edges, and the ecosystem around them takes time to settle.
The lock-in questions to ask
A unified framework is convenient precisely because it nudges you deeper into one vendor’s world. That is the trade you are making, and it is worth making with your eyes open. Before you standardise, get straight answers to a few questions:
- Model portability — can you point the framework at non-Microsoft models, or are you steered firmly towards one provider’s hosted endpoints?
- Data residency — where does your data go when an agent runs, and does that satisfy your obligations to UK clients?
- Exit cost — if you wanted to move to a different framework in eighteen months, how much of your agent logic is genuinely portable versus glued to this one?
- Open governance — how much of the roadmap is community-driven versus set entirely by the vendor?
None of these is a reason to walk away. They are simply the difference between adopting a tool deliberately and drifting into a dependency you didn’t choose.
What this means for a small UK team
For a professional-services firm, the practical read is reassuring but not urgent. Consolidation lowers the risk of betting on the wrong horse, and GA means you can plan around the framework rather than chasing a preview. If your business already runs on Microsoft tooling — Office, Azure, the wider stack — the unified Agent Framework will feel like a natural extension and the integration friction will be low.
The disciplined move is to run one small, well-scoped pilot — a single internal workflow, with a human reviewing the output — and use it to test both the framework and your own appetite for the lock-in trade. Standardise once the pilot proves out, not because the GA badge looks tidy. Fewer abstractions is a genuine win; just make sure you’re choosing the stack, rather than letting the stack choose you.
