CI for the agent era
Your new buyer is an AI coding agent.
Can it actually adopt you?
Agentboarding turns real, sandboxed Claude Code agents loose on your dev tool — they try to install & integrate it from your public docs alone. We record the run, replay the WebFetch → Haiku pipe to prove where it broke, auto-fix it, and re-run a fresh agent to prove the gap closed.
Not "are you cited?""can an agent succeed at using you — and where does it break?"
▶ Run the interactive replay ▶ Raw CLI recording (real boot)From Rozz's own logs: on Mar 27 a real Claude Code session fetched genymotion's pricing Q&A at 18:30:56 and the gmsaas runbook 10 seconds later — evaluation → implementation, no browser. 12 of 14 Claude-User requests came from Claude Code.
Interactive replay · driven, not watched
Pick an arc. Step through a real recorded run.
Left: the agent's terminal. Right: the trajectory + the AX Score animating red→green on the re-run. Use the controls to drive it.
The money moment — what the agent actually read
Claim survival — real Haiku, same page
The loop, end to end
run → fail → autopsy → root-cause → auto-fix → re-run green
adb shell · AX 0.96 · 0 instances leakedThe raw CLI · unedited
AI site vs the real site — same agent, same goal
Same agent (Opus 4.8 + Claude Code), same goal — adopt gmsaas and boot a device. The only thing that changes is what it reads. Both succeed; the AI site removes the friction the agent otherwise has to overcome itself.
| Same agent · same goal | Real genymotion.com (no AI site) |
AI site (Rozz) |
|---|---|---|
| AX score — executed adoption outcome | 0.85 | 0.96 |
| Pages the agent had to read | 6 doc pages | 1 runbook |
| Adoption artifact | agent wrote its own runbook | used the ready-made runbook |
| Key facts survive WebFetch→Haiku | ~0.0 | 0.60 |
| Auth-token leak | none — used the env var | none — and the page warns against it |
| Reached a working integration | yes — the hard way | yes — pricing→runbook in ~10s (real Mar-27 session) |
Three recorded runs below back every cell: the AI-site boot (0.96), the detect-&-fix loop (0.35→0.94), and the real-site control (0.85).
Why this is new
Everyone measures presence. We measure — and repair — whether an agent can succeed.
vs GEO / AI-visibility
Profound, Scrunch, Peec count whether you're mentioned. They never execute. We run the integration and ship the fix.
vs coding-agent evals
SWE-bench / SWE Atlas grade the agent. We grade your product's adoptability by agents — the opposite subject.
vs agent infra / docs
Composio wires agents to APIs; Mintlify renders docs. None watch a third-party agent try and fail to adopt you.
The wedge: Agent Experience (AX) testing — Lighthouse + Sentry, but the user is the autonomous agent that's now your buyer. Claude Code, grading Claude Code, and shipping the fix.
How this demo works
This is a replay — Agentboarding is an agent-orchestration CLI
A live "click-to-boot" button would expose API keys and bill cloud minutes to anonymous visitors, so this page replays a real recorded trajectory. The tool spawns a real headless claude -p agent, runs the WebFetch→Haiku replay out-of-band, classifies the failure, auto-fixes, and re-runs — with a hard before/after instance teardown so a run can never leak a billable device.
The real run, verbatim
Built & verified
Deterministic test suite green; WebFetch→Haiku fidelity validated against real Haiku; the live device boot recorded as committed evidence. Built for Claude Build Day with Opus 4.8 + Claude Code.
Recorded replay of a real run, 2026-06-13. Real agent, real Genymotion cloud device (Android 14), real teardown (0 instances leaked). The Haiku summaries shown are verbatim real outputs. In Arc B the funding wall is reproduced via
--funding blocked (real Genymotion error signature, injected) and the x402 settle is test-mode — no funds move; the device boot is real.