Topic

AI Agents

4 min read

Generate code from the spec, not the spec from the code

Agents guess field names when they never see the contract. Serve the spec over MCP, generate one verified vertical slice at a time, run scenarios against localhost, and rescan to catch drift.

Read article
4 min read

MCP vs function calling vs ChatGPT plugins: what API teams actually need in 2026

Function calling is a model capability; plugins were a ChatGPT-era distribution experiment; MCP is a transport and discovery protocol between clients and tool servers. They are not competitors, and confusing them produces bad integrations. Here is the layered view and where your OpenAPI spec fits.

Read article
2 min read

We gave a partner’s AI agent 12 of our 47 endpoints. Here is how.

A concrete Cloud walkthrough: publish a managed Streamable HTTP MCP endpoint from one spec, expose twelve operations out of forty-seven, gate agents with a rotated bearer key, inject upstream auth, and pin the partner to an immutable version.

Read article
4 min read

Stop pasting API docs into AI coding agents: serve the spec over MCP

Pasted docs rot mid-session, burn tokens on every conversation, and leave the agent guessing field names. An MCP server generated from the OpenAPI spec gives an agent callable tools with typed inputs instead of prose. Here is when that wins and when it does not.

Read article
2 min read

Your API already describes the tools your AI agent needs

Hand-writing tool wrappers around existing order and inventory endpoints means maintaining every contract twice. Generate the MCP server from the OpenAPI spec, start with read-only tools, and widen the surface deliberately.

Read article
2 min read

A 200 is not done: make the AI walk the business scenario

The create call returned 200, the detail view 404s, and cancel accepts an illegal state. Scenario tests chain real calls, pass data between steps, and hand the agent evidence to fix with.

Read article
2 min read

Don't block frontend work on a backend that isn't ready

Fake objects in components demo well and integrate badly. Agree on the coupon validation contract first, then let a spec-driven mock answer real HTTP while the service is still being built.

Read article
2 min read

Your AI codes fast. The integration is still wrong.

Generated screens fail on limit versus pageSize, data.records versus data.items, cents versus dollars. The fix is not a better prompt; it is letting the agent query the real contract and verify against real responses.

Read article
3 min read

When AI writes your system, who defines what done means?

A booking app built from one prompt can pass every visual check and still fail on time zones, double bookings, and stale cancel responses. Teams need a contract both people and tools can query, change, and verify.

Read article