Talk to the agent about a live copy of your app.
A chat where the agent works for you against a fresh environment it stands up when the conversation needs one. Ask for an exploration, a reproduction, a new test, a suite run or a filed issue, and it does that thing with the same tools and evidence a scheduled run has.
userSet up a housing event from the captured Expo dump and check whether a hotel with zero inventory can still be booked.
toollist_vendor_dumps() → [{ code: "expo", name: "Expo 2026", features: ["housing","registration"] }]
configure_sut({ dump: "expo" }) → environment requested; ready in ~6 min
agentEnvironment is up with the Expo event restored. Signing in as the organizer with the catalog login test, then opening Housing → Hotels.
toolbrowser_navigate("/organizer/…/housing/hotels")
browser_click("Grand Plaza") · set inventory 0 · publish
agentInventory is 0 and the hotel still appears as bookable on the participant flow; I completed a booking for it. That looks like a defect. Do you want a repro test and a filed issue?
userYes, both.
toolcreate_test({ name: "housing-zero-inventory-still-bookable", … }) → disabled, pending review
file_issue({ title: "Hotel with zero inventory remains bookable", … }) → #1804, watched
The environment comes when it is needed.
Nothing boots when you open the chat. The agent can list the project's captured data sets from the first turn and requests an environment configured with one when the task calls for it. While the box boots the conversation waits; tool calls stream into the transcript as they happen.
Everything the agent produces in a chat is a first-class object: a test lands in the catalog (disabled), an issue is filed and watched, a run appears on the Jobs view with its evidence.
- Per-project caps on concurrent sessions, an idle window and a lifetime ceiling; the run timeout does not cut a live conversation short.
- One implementation, every harness. The chat backend is shared by the built-in loop and Hermes; a fix reaches both.
- Evidence at every turn. Screenshots publish at turn boundaries, not at a session end that may never come.
Where it stops today
- A session holds an environment while it is active, which costs compute; the idle and lifetime ceilings exist for that reason and end the session when they hit.
- The agent's tools in chat are the same as in a run. It cannot change project settings, merge anything or write to your repository from a conversation.
- Only one project's environment per session; switching products means a new chat.
- Sessions reconnect if the underlying job dies, but a long silence is still a long silence: watch the environment status line, not the typing indicator.