Define the interface contract before you buy any AI agent framework

OpenAI, Anthropic and Google ADK all point to the same operator risk: teams buy agent frameworks before deciding who approves tool calls, what evidence is kept, and where work rolls back.