I recently spent some time digging into Anthropic's Commerce Agents to see how their shopper and merchant assistants actually perform under the hood. I ran 42 different test scenarios across four mock storefronts (retail, travel, telecom, and entertainment) to understand how these agents handle product discovery, budget constraints, and backend business logic. If you are building AI into your commerce flows, here are the main takeaways you need to know: 1. Chat doesn't override system controls: Just telling the merchant AI "I approve this update" in the chat doesn't bypass the application's actual approval gates. The separation of powers remains intact: agents can prepare and propose changes, but your application logic still controls the final execution. 2. Juggling multiple constraints is still hard: When given a strict budget and a specific list of required items, the AI sometimes struggled to satisfy everything at once. In one retail test, it hit the budget perfectly... but quietly left out essential camping gear, treating those items as "optional." 3. Mind the backend disconnect: What the agent promises in chat isn't always what the backend executes. I saw discrepancies where the AI's quoted travel price didn't match the final cart total, and marketing campaign drafts dropped dates that the AI had clearly supplied. The Bottom Line: When implementing commerce agents, your top priority must be validating the resulting state. Don't just trust the chat output—programmatically verify that the final cart or proposal actually meets all of the user's explicit constraints before moving to checkout. Curious to dive deeper into the architecture, tool behaviors, and where the boundaries lie? Check out my full write-up: redacted