How to Make a Web Service Usable by AI Agents

Designing a web service that AI agents can effectively interact with requires focusing on two key aspects: accurate state representation and clear, well-defined actions. Websites, marketplaces, and cloud applications still exist, but they need to be integrated in a way that allows an agent to perform tasks rather than merely render screens.

Consider a service for ordering desks. The user specifies size and delivery before a certain date. The service must return factual information about availability, current price, delivery options, and any limits. The text used to describe the product should be separated from the state used to make the purchase. OpenAI's Agentic Checkout Spec emphasizes the importance of a rich cart state, including items, pricing, taxes, shipping, discounts, totals, and status. The merchant's response, not the model's memory of an earlier page, should be the authority.

When responding, ensure the information is explicit. If an item is out of stock, do not assume the caller will infer this from missing buttons or empty fields. If shipping information is unavailable, return this uncertainty instead of providing an estimated date. Define small actions with a clear contract. Creating a checkout session is different from completing it, and retrieving a customer record is different from modifying that record. MCP provides a standard connection between AI applications and data sources or tools.

The service still needs its own validation, access checks, and clear operation boundaries. Agents should be treated as another client of the same business rules, not as a shortcut around them. If a complete-order request is lost, retrying blindly could result in a second order. Use idempotency, request tracing, authenticated requests, and safe retries. OpenAI's checkout specification explicitly calls for these measures.

Identity validation is crucial. Do not trust a caller simply because it identifies itself as an agent. Validate its identity and the account it can access, separating permission to read from permission to change or buy. Google's UCP guide and OpenAI's commerce docs provide direction, but access and feature scope remain important.

OpenAI limits Instant Checkout to approved partners, and Google's guide has a waitlist and roadmap. The useful promise is that this operation can be discovered, called within its limits, and checked after it runs.

The ultimate goal is to have one verified workflow before expanding to a broader agent platform. After integrating one service, define the success state and exercise failure cases like stale availability, wrong account, lost responses, and changed totals. Always provide a human route when the agent route fails. This approach ensures that a website becomes part of an agent-operated internet, with smaller, more manageable operations that are discoverable, callable, and verifiable.