Developer tools have always had one buyer: the developer. Everything you built, from docs to pricing to DevRel, was built for them.
But there's a second one now: the coding agent. It doesn't sign the contract. But it finds the options, tries them and picks one, and the developer usually ships what it picked.
It's already here. On OpenRouter alone, coding agents made about 60 million model requests in a single day.
That puts the agent right in the awareness layer of your funnel. Before a developer has heard of you, their agent may already have tried you, and every try has a cost.
Give Claude Code the same job twice: a payment link for a $29/month subscription, once with Stripe and once with another billing API. You get a working link both times. But what did it take to get there?
We measure that path on Devtool Arena, our public benchmark. Every API in a category gets the same setup: one written task, an API key, a clean sandbox and no other human help. Then we record everything Claude Code does on the way to working code: every tool call (each page it fetched, command it ran or file it wrote), every error and every token.
With Stripe, the agent used about 150,000 tokens, 6 tool calls and 2 errors, and had a working link in 37 seconds. Cost: 8 cents.
With the other API, it used 2.8 million tokens, 58 tool calls and 11 errors, and took almost six minutes. Cost: $1.09.
Both are serious companies. Both have docs, a typed SDK, llms.txt and an MCP server. On a checklist, they look the same. The gap only appears when an agent tries them.
Developers now find tools through their agent
You used to find developer tools the slow way: a search, a thread, word-of-mouth recommendation. Your SEO, content and DevRel were built to show up in exactly those places.
Now you skip that step, and so do your developers. You ask Claude Code for "a PDF parser" or "auth for this Next.js app", and the agent decides which tools to suggest, from what it already knows and what it can fetch. Claude Code, Codex and Cursor decide which names come up before you've heard of any of them.
That makes the agent the first gatekeeper of your product. If it doesn't know you, or can't find your docs, you never make the list. And whatever it concludes, it passes on to whoever asked.
What the agent judges you on
The agent doesn't care about your positioning. It cares about four things:
- Can it find the right docs? Not your marketing site: the page that tells it what to install and which endpoint to call.
- Can it install and authenticate? Keys, scopes, redirect URIs, environment variables.
- Can it finish the task? Working code, not code that looks right.
- Can it recover when something breaks? A clear error with a next step, or a guessing game through dead links.
Think of it as a review, written in tool calls. This is the new metric.
Most products aren't ready for this buyer
Across the 65 APIs in our Q1 2026 report, in the best case, exactly one let the agent finish with no human help at all. On the other end, there were 16 where the agent kept failing.
The other 48 sit somewhere in the middle: mostly working, or needing a human to step in.
The biggest gaps are between products that compete directly, and that's the part that should worry you. Take two auth APIs from the benchmark. One got the agent to working code in 4 tool calls with zero errors. The other took 61 tool calls and 23 errors, most of them on redirect URIs and middleware. When a developer's agent can pick either one, that run decides who gets the integration.
You lose before anyone hears your name
If your product loses at this new game, nothing tells you, because it happens before the developer has heard of you. No one searches your name, no one lands on your homepage, no one types you into "how did you hear about us?". The developer gets a diff that works, with someone else's SDK in it, and never learns you were an option.
Your awareness metrics were built to count people noticing you. Now the first thing to notice you is an agent, and almost nobody can see what it decided. The only way to know how you did is to run the integration and read the bill.
Try this today: give Claude Code your own quickstart and a fresh API key, and count the tool calls it takes to get working code.
