Your Lambdas were built for callers you could reason about: an API Gateway with a rate limit, a queue with a batch size. Now one of them decides for itself how often to invoke you, with arguments inferred from a docstring you never treated as production config.

One question becomes two hundred calls, a six-step agent chain compounds into eighteen sequential invocations inside one user request, and idempotency keys stop matching because a model doesn't retry, it re-reasons. This session takes the tool provider's chair: what to harden in functions you already run, how to architect them and more.