AutoGen's RoundRobinGroupChat or SelectorGroupChat run until a termination condition fires. If your biller agent issues the payment request but the termination string never appears in the chat (typo, model paraphrasing), the group keeps looping. Use TextMentionTermination with a specific tag like PAYMENT_REQUESTED that you instruct the biller to output verbatim. Validate this works with a non-paying test before deploying.

AutoGen tools registered to one agent are not automatically available to other agents in the team. If the researcher tries to call payment_request_tool, the call fails because the tool is scoped to the biller. This is usually correct, but first integrators sometimes assume tools are team-wide. They are not; assign tools per agent.

AutoGen 0.4+ is async-first. payment_request_tool is awaitable; calling it as a sync function from a non-async context returns a coroutine, not a result. Use within asyncio.run() at the top level, or ensure your enclosing function is async. The error 'coroutine was never awaited' usually means you called the tool from a sync function.

Each agent in AutoGen takes a model_client. If you instantiate one OpenAIChatCompletionClient per agent (rather than sharing one), you pay 2x or more on token usage because each client maintains its own context. Share a single model_client instance across the team unless you have a specific reason for separate config (different temperature, different model per role).

AutoGen group-chat runs are typically short-lived (they end at the termination condition). When the payment.confirmed webhook fires minutes later, your original team.run() call has already returned. You need an async job queue or persistent task store (Redis, Postgres) to record the payment-pending state and trigger a new team.run() that picks up where the first one left off. Do not try to keep the original team alive waiting; AutoGen is not designed for long-lived runs.