AutoGen

AutoGen was Microsoft's first serious attempt at multi-agent orchestration, and it showed real promise before being deprecated in favour of Agent Framework. The core idea remains sound: define agents with distinct roles, let them converse to solve problems, and watch emergent behaviour handle complexity you didn't explicitly programme. I used it for a research assistant pipeline where one agent scraped papers, another summarised findings, and a third synthesised conclusions. The conversational handoffs worked surprisingly well when the task had clear boundaries. The autonomy here is conversational, not environmental. Agents don't browse the web or manipulate files without your plumbing. Instead, they debate, critique each other's outputs, and iterate towards a solution. This shines in code review workflows: one agent proposes changes, another flags edge cases, a third runs tests. The back-and-forth catches mistakes a single LLM pass would miss. Failure modes are predictable. Agents can loop endlessly if termination conditions are vague. Token costs spiral when conversations run long. The framework assumes you understand prompt engineering well enough to define coherent roles. If your agent personas are mushy, the conversations devolve into agreement theatre where nothing meaningful happens. Compared to CrewAI, AutoGen feels lower-level and more research-oriented. CrewAI gives you opinionated patterns for common tasks; AutoGen gives you primitives and expects you to build your own patterns. LangGraph offers finer control over state machines but requires more boilerplate. AutoGen sits in the middle: flexible enough for experimentation, structured enough that you're not writing raw API calls. The deprecation is the elephant in the room. Microsoft has moved active development to Agent Framework, which shares DNA but isn't a drop-in replacement. If you're starting fresh, Agent Framework is the safer bet. AutoGen still works and the community remains active, but you're building on a sunset platform. For research projects or learning multi-agent patterns, it's still valuable. For production systems, the migration risk is real. Documentation is thorough but assumes familiarity with agent concepts. The examples are academic-flavoured, which helps if you're exploring ideas but less so if you need to ship a feature by Friday.