Originally published on the Postman blog.
Agent-ready step one: inventory your APIs and agents, then map them.
If you build backend or platform services, your next heavy API consumer probably isn't a person. It's an agent that can't read your README, can't ping a teammate when a call returns a vague 400, and can't guess which of three similar endpoints is the right one. The 2026 State of the API Report puts numbers on that gap. This post walks through the three fixes that close it, in the order we'd do them.
The gap
One in four organizations now run AI agents that use APIs in production. Only 13.2% say their APIs are fully ready for them.
Picture a team that shipped an agent last quarter. It books travel, issues refunds and sends the weekly report, and to do that it calls a dozen internal APIs. Those APIs were built years ago for developers. A person could read the docs, ask someone when a call failed, and work around the odd error. An agent can't. It needs clear descriptions, predictable errors, tight permissions, and a way to find the right endpoint in the first place.
Most APIs still aren't designed for that caller: 38.7% of teams still design APIs primarily for human developers reading docs, and only 5.6% design primarily for agents.
API-first teams are further along
Fully API-first teams are 2.6x as likely as non-API-first teams to run agents in production: 43.2% vs 16.7%.
One caveat: this is an association, not proof of cause. Teams that already invest in APIs may simply be the ones that invest in everything. The pattern is still worth noticing.
MCP is real. Discovery isn't solved.
MCP has moved quickly into real use: 55.6% are in production or already experimenting with it, and 23.8% are already in production.
But agents can only use what they can find. 37.7% have not enabled any discovery for agents, even as 12.2% say recommendations from AI agents or coding tools are already a top factor in choosing which APIs to use. If an agent can't find your API, it will use someone else's.
Agents write, and credentials lag
In Postman's own platform data (not the survey), agents mostly change things rather than just read them: 51% of tool calls through the Postman MCP server create or change collections and requests, versus 46% that only read.
Meanwhile, 22.1% still store credentials in environment variables or .env files, and only 13.4% use short-lived credentials. A caller that writes, holding a long-lived secret from a dotfile, is the pattern to fix first.
The three fixes, in order
1. Build the map and make it discoverable
Start with an inventory of your APIs and agents, then map it with the Context Graph API so people and agents can both find what exists. Even a plain file in a repo works as a first pass. Here's an illustrative shape:
apis: - name: refunds-service owner: payments-platform spec: specs/refunds.openapi.yaml consumers: [support-agent, billing-ui] writes: true agents: - name: support-agent calls: [refunds-service, orders-service] credential: short-lived # flag anything still in .env The point is a single list that answers: what exists, who owns it, which agents call it, and which of those calls write.
2. Generate docs from the spec
Agents need descriptions that match what the API actually does today. Generate docs from your API definition so they stay current for people and agents, instead of relying on hand-edited pages that drift.
3. Put identity and a gateway in front of agents before you scale them
Treat agents as non-human callers with their own identity and short-lived credentials, and route their traffic through a gateway so there is a control point where policy applies.
Read the data
Read the full post on the Postman blog, then open the 2026 State of the API Report for the data by role, region and industry.
What does your agent inventory look like today? I'd like to hear what you're seeing in your own stack in the comments.





