Agent experience (AX) is how easily AI agents can find, understand, and use your product. For AI, gaming, and Web3 software teams, we run coding agents, browser agents, and assistants through your docs, signup, and API, log where they get confused or quit, and fix it.
Five agents get the same task and a clean start. Pick a task to see where each run stalls or stops, and what fixes it.
| Agent | 01Find signup | 02Create account | 03Verify email | 04Create key | 05Store key |
|---|---|---|---|---|---|
| Terminal coding agent | Cleared | Cleared | Stopped | Not reached | Not reached |
| IDE coding agent | Cleared | Cleared | Stopped | Not reached | Not reached |
| Browser agent | Cleared | Asked | Cleared | Cleared | Cleared |
| Chat assistant | Cleared | Stopped | Not reached | Not reached | Not reached |
| Computer-use agent | Cleared | Cleared | Cleared | Asked | Cleared |
A CAPTCHA ended the chat assistant's run. The browser agent asked whether the free tier needs a card.
FixLet signed agents past the CAPTCHA, and answer the card question on the form.
Both coding agents stopped at the verification email. A terminal can't open an inbox.
FixA CLI login that issues a test key in the terminal.
The computer-use agent asked which key scope the quickstart needs.
FixName the scope in the quickstart and preselect it.
Coding agents such as Claude Code, Codex, and Cursor, plus browser agents and ChatGPT agent, attempt the jobs your users hand them, several runs each.
Asked to solve a problem in your category, does an agent pick you, a competitor, or its own code?
Markdown pages, an llms.txt map, stable URLs, and code samples that run as written.
Email verification, CAPTCHAs, card walls, and the bot rules that block the agents you want as users.
Output an agent can parse, errors that say what to do next, and MCP servers with fewer, clearer tools.
Every question an agent stops to ask, with the page where the answer belongs.
Write the tasks your users hand to agents, then run each agent on each task several times in our own RL environments for a baseline.
Rank failures by how many runs they stop, and agree the fixes in a written scope.
Fix docs, errors, and signup in your repository, or hand your team a ranked fix list.
Re-run the same tasks against the baseline, and again after releases and new models.
Illustrative audits, not client work. Each names what we would run, find, fix, and measure.
More signups come from coding agents every month, and few of those accounts ever make a successful call.
Run
Three coding agents asked to add the API to a sample app from an empty repository.
Find
Agents stall on email verification and guess the auth header.
Fix
A CLI login that issues a test key, and an error that names the missing header.
Measure
Runs that reach a working call, and steps and tokens per run.
Developers ask their coding agent to add wallet login, and it picks a rival SDK or an old version of yours.
Run
Agents asked to add wallet login to a Next.js app, with and without naming your SDK.
Find
Old package names still rank, and the testnet faucet sits behind a CAPTCHA.
Fix
Deprecation notices in the registry, and a faucet path agents can use.
Measure
How often agents choose you unprompted, and runs that reach a signed testnet transaction.
Studio engineers prototype with coding agents, and the agents rarely suggest this backend for matchmaking.
Run
Agents asked to add matchmaking to a Unity and a Godot project, then to compare backends.
Find
No page names the supported engines, and the pricing calculator is empty without JavaScript.
Fix
One page per engine with a working sample, and pricing as a plain table.
Measure
How often agents recommend you, and runs that reach a working match.
| Deliverable | What it is |
|---|---|
| Run transcripts | Every run, step by step, with where it stalled and why. |
| Failure map | Each failure by step, with how many runs it stopped. |
| Question log | Every question an agent asked, with the page that should answer it. |
| Ranked fix list | Fixes ordered by runs recovered, with each change written out. |
| Re-run report | The same tasks after the fixes, against the baseline. |
In July 2026, AI agents made 66% of the measured traffic on docs hosted by Mintlify, up from 15.2% in January. When they fail, they leave without telling you why.
Tell us the task you want agents to finish. You get a written scope, with every deliverable and price, before any work starts.