According to @AnthropicAI, autonomy in AI systems is co-constructed by the model, user, and product, meaning pre-deployment evaluations alone cannot fully characterize real-world behavior; as reported by Anthropic’s blog linked in the tweet, the company advises developers to test autonomy across product contexts (e.g., UI constraints, tool access, and guardrails), monitor post-deployment behavior with red-teaming-in-the-wild, and design incentives that reduce unintended persistent agentic behavior. According to Anthropic, policymakers should calibrate oversight to deployment context, require evidence of post-deployment monitoring, and prioritize incident reporting standards that capture product-mediated autonomy. As reported by Anthropic, these recommendations aim to improve model governance, reduce emergent risky behaviors when tools and memory are enabled, and align enterprise risk management with real user interactions and product design choices.