Atlas for AI Agents
Describe the test. Let the agent build it.
Load testing has always had a gap between wanting a test and having one. Someone has to click through the flow, capture the requests, work out which values are per-session tokens, wire up the data files, size the run, and then interpret a report full of percentiles. It is a day of work for a thirty-second question: will this hold up on launch day?
Atlas closes that gap. Every part of it — recording, test plans, data files, machines, runs, results — is reachable through one command-line tool that speaks JSON, and it ships with the guidance an AI coding assistant needs to use it properly. Point your assistant at Atlas and ask in plain language:
Record a checkout on our staging site, then run it with 500 users for ten minutes and tell me where it breaks.
It records the flow in a real browser, turns the recording into a genuine load test, checks the plan will actually report, smoke-runs it, asks you to confirm the real run, and comes back with what broke and at what point.
New here? Set it up in two commands, then see what it can do.
Why this works with Atlas
One tool, every surface. The agent does not need infrastructure credentials, a database, or your cloud account. Everything goes through the same CLI you use, so your permissions, your account's quota and your machine limits apply exactly as they do in the browser.
Machine-readable by design. Every command takes --json. Failures print JSON on standard error, so clean output stays clean. atlas schema --json describes the entire command tree, so an agent can discover the surface instead of guessing at flags.
It knows the traps. Load testing fails quietly. A data file that was never uploaded silently removes an entire thread group — with zero errors to explain it. A plan whose structure defeats reporting runs perfectly and tells you nothing. Fast responses that are really login pages look like excellent performance. The bundled guidance teaches an assistant to recognise all of these before you act on a number that means nothing.
The documentation travels with it. atlas docs prints this documentation offline, so an agent in a sandbox or a CI container still has the manual.
What you get
| Recording without clicking | The agent opens a real browser, reads the page as an accessibility tree, clicks through your flow and names each step — or hands the window to you for the parts only you can do, like a password or an MFA prompt. |
| A real test, not a replay | It spots the values every simulated user would otherwise repeat — credentials, search terms, product ids — and moves them into a data file, so you are testing your application rather than your cache. |
| Plans that actually report | Before anything runs, atlas plan check catches the three failures that produce an empty dashboard an hour later. |
| Sizing you can defend | It reads your real capacity and explains what 500 users means in machines, because each machine runs the whole plan. |
| Results in sentences | Not just p95 numbers: which step degraded, at what user count, and what the errors actually say. |
| CI that fails honestly | Pass/fail thresholds, JUnit output and exit codes, written into your pipeline by the same assistant. |
Guardrails
An agent that can start a load test can spend money and put load on production. The bundled guidance is explicit about this, and a good assistant will:
- Confirm before any real run. Machines are powered on for a test and billed to your account. Smoke tests are cheap and encouraged; a 20-machine run is not something to start on an assumption.
- Never stop a run it did not start. Someone else's test is someone else's.
- Ask rather than guess about the things that change the result: which project, how many users, how long, ramp-up, fixed machines or a cluster.
- Report what happened, not what should have. Including failed thresholds, a thread group that vanished, and its own uncertainty.
- Leave your files as it found them. It downloads before it edits, and it restores anything it changed for debugging — think time included.
Everything an agent can do, you can do by hand with the same commands. Nothing here is required, and nothing here is hidden from you.
Which assistants
| Assistant | How it reads the guidance |
|---|---|
| Claude Code | Skills — loaded automatically when your request is about load testing |
| GitHub Copilot, Cursor, Codex, Gemini CLI, Windsurf | AGENTS.md in your project root |
| Anything else | atlas schema --json, atlas docs, and --json on every command |