The three loop topologies (ReAct, plan-and-execute, orchestrator/subagent), why your agent goes in circles, and what Runner.run is actually doing for you.
Why agents forget, the three layers of memory you actually need, and how to wire a vector store into the loop without the framework hiding what's happening.
What an agent tool actually is, how OpenAI Agents SDK, Anthropic, and Google ADK each declare one, and why every tool is a blast-radius decision.
