AG2

AG2 kept the conversation model and cleaned up the ergonomics around it. The trait you feel every day is tool calling by decorator: you write an ordinary Python function, annotate it, and an agent can call it, so your logic stays in code you can test instead of in a prompt you can only reread.

Pick it if your team already thinks in Python and wants several agents talking with a person able to interrupt them. Pick AutoGen if you want the original and the lab that publishes behind it, because AG2 is a community project carrying a much older name.

The risk is the split. Version 1.0 breaks compatibility with the pre-1.0 code, and the earlier framework keeps shipping as ag2-classic. A team with running agents must port or freeze. Every tutorial written before the change describes whichever package the reader did not install, and there is no way to tell from a code sample which one it targets.

Budget the port before adopting and pin the package name in requirements. What it does right: the maintainers say plainly that v1.0 is not a drop-in upgrade instead of pretending the rename was cosmetic.

1. Context arrives from two places: the message history of the conversation itself, and pluggable knowledge stores the framework queries on an agent's behalf. 2. Planning is emergent. The plan is whatever the agents say to each other, which makes a transcript readable and an outcome hard to bound. 3. Actions are registered Python callables. 4. Verification is a human-in-the-loop step, not a test run.

No benchmark appears in the repository, so there is nothing to audit. That is preferable to a self-scaffolded number, and it leaves a reader with no evidence of capability beyond the examples.

The asset here is a name and a mailing list. Everything is bring-your-own-key with no paid tier, so there is no revenue line and no pricing power to test. Adoption sits near 4,900 stars, respectable for a Python library and thin for an organisation that will eventually need a funding story.

Moat: none a competing framework cannot copy in a quarter. The routes out are a hosted control plane, a support contract, or absorption by a platform that wants the orchestration layer and the recognition attached to the name. Position: adopt the code, do not build a vendor relationship around it.

The demo is agents talking to each other with a person able to step in. Procurement is short because there is nothing to purchase: the cost is engineer time. It runs headless, so it fits the pipelines we already operate. There is no single sign-on because there is no console, and no audit trail beyond logging we write ourselves.

Onboarding is about a week for a mid-level Python engineer and impossible for the half of the organisation that writes TypeScript, because the framework is Python only. Approved with conditions: one team, one service, our own logging around it.

Apache-2.0 and plain Python, so I can read the loop and patch it in place. The install takes extras, pip install ag2[openai], and the model layer accepts Ollama, so the whole thing runs on my hardware with no vendor key anywhere in the path.

The gap is protocol. It is not an MCP client and not an MCP server, so every tool I already run behind MCP has to be rewritten as a decorated function before these agents can reach it. That is a wrapper layer I did not want to own. Forking is trivial and the licence is honest. I would still rather they shipped a client.