TestStudios — Testing becomes a conversation

90-second brief

How to read this

The low-code IDE still assumes someone will record, edit, and maintain a suite. Non-technical testers, PMs, and support still bounce off a script-shaped product. Web and mobile still live in different tools. I needed a Studios that removes the scripting layer, not another nicer editor for it.

• Keep low-code as the only Studios

  Died

  Record-and-heal is a real product. It is still a barrier. We would keep selling to SDETs and losing the people who file the bugs.

• Fully autonomous agent, no plan

  Died

  An agent that clicks before a human sees the plan is a liability in QA. Wrong path, wrong env, burned credentials. Fast demo. Unshippable for a release gate.

• Chat-first, Plan mode, HITL

  Shipped

  Conversation is the authoring surface. The agent proposes a plan, waits, then drives a real browser or device. Pause and take over stay first-class. This is what shipped in 0.1.0.

Decision

Remove the scripting layer. Testing is a conversation with a plan you can approve. Web and mobile share one desktop app. Memory of the product is the category claim — not a one-off chat that forgets.

Result I will defend

Shipped on the skills-first stack: English in, structured bug report out. Chat, Plan mode, live view, OTP. Web and mobile in one macOS app. Weekly authors are not on a dashboard — I will not invent a seat count.

What I got wrong

I almost treated this as a skin on the low-code IDE. It is a different ICP. If the first screen is still a test table, the no-code bet is already lost. I also cannot pretend cloud device emulation is a device lab — Appetize is how we ship mobile without hardware. That constraint has to stay visible.

Who loses if this is wrong

The release manager, if an unreviewed agent path papers a defect. The tester, if chat cannot finish login. The SDET, if no-code floods suites nobody can debug.

Metric spine

North star

Time-to-first-executed-test

A PM or QA can describe a flow and see the agent driving a real browser or device without writing a locator. If we lose to “just ask ChatGPT and click along,” we lose the wedge.

Guardrail

Unreviewed action / silent drive

The agent must not burn a production login or a paid environment because Plan mode was skipped. Pause-before-act is a product requirement, not a preference.

Not instrumented

Weekly active authors who never open a script

That is the ICP proof. 0.1.0 does not have a published cohort. I will not invent one.

Not instrumented

Plan-accept rate vs takeover rate

Would show whether Plan mode is load-bearing or theatre. Not on a dashboard yet.

Not instrumented

Web vs mobile run mix

The category claim is both platforms in one tool. Split is not instrumented.