Work / SuperSquad — autonomous QA with evidence

SuperSquad — autonomous QA with evidence

SuperSquad turns plain-English product intent into scheduled browser QA with reproducible, evidence-backed findings. In progress.

Problem

Web teams need reproducible QA findings from plain-English product intent.

Constraints

A confirmed finding requires reproduction steps and linked evidence; incomplete runs must remain visibly incomplete.

What I did

SuperSquad is an autonomous QA product I'm building for web teams. A team describes what matters in plain English and gets scheduled browser-based runs that produce evidence-backed findings.

It is not a chatbot and not a Playwright recorder. It is a QA control plane: translate intent into coverage, execute journeys, capture artifacts, validate findings, and publish reports that a stakeholder can understand and an engineer can reproduce.

Product invariant

The core rule is strict:

No confirmed finding without reproducible steps and linked evidence.

This rules out a large class of attractive but operationally weak outputs. A model’s description of a broken page is not enough. A confirmed finding needs the expected behavior, observed behavior, reproduction path, and evidence such as screenshots, recordings, console logs, network traces, DOM state, or generated test context.

Two state owners

The product separates organization state from execution state.

An organization state owner manages applications, environments, product profiles, coverage maps, schedules, entitlements, usage, and accepted terminal results. A run owner executes one immutable snapshot and owns its browser session, retries, evidence metadata, finding candidates, and terminal proposal.

flowchart LR
  Command["Command center"] --> Organization["Organization state"]
  Organization --> Snapshot["Immutable run snapshot"]
  Snapshot --> Run["Run execution owner"]
  Run --> Browser["Browser session"]
  Run --> Evidence["Evidence in R2"]
  Run --> Proposal["Result proposal"]
  Proposal --> Organization

The run can be long-running and failure-prone without turning the organization object into a synchronous side-effect hub.

Immutable execution

At admission time, the organization packages the application, environment, product intent, coverage, allowed domains, and blocked actions into a snapshot. The run executes that snapshot rather than reading mutable configuration throughout the run.

That makes reports explainable. The report describes a known version of the test intent, even if the application profile or schedule changed while the browser was running.

Evidence as a first-class artifact

Screenshots, recordings, console output, network traces, generated tests, and reports are stored in R2. D1 stores organization, authorization, usage, and product records. Queues and Workflows handle schedules, retries, and long-running orchestration. AI Gateway keeps model routing behind a controlled boundary.

The browser is a capability inside the system, not the product itself. The product is the loop that turns intent into a reproducible claim.

Guardrails

Runs need explicit limits around allowed domains, credentials, destructive actions, captchas, email verification, payment flows, browser time, model usage, and isolated probes. The system must report an incomplete or blocked run rather than silently claim success when it could not test the requested behavior.

SuperSquad is in progress under Developer Labs. It is the clearest expression of the systems I like building: agents with state, tools, evidence, contracts, and an explicit refusal path.

Result

In progress. The architecture separates organization state, execution state and evidence storage.

Tradeoffs

Long-running browser execution needs isolation, limits and an explicit incomplete outcome. Product intent and run evidence must remain tied to the same immutable snapshot.

Stack

Cloudflare WorkersDurable ObjectsR2D1Playwright MCPAI GatewayTanStack Start

Explore more problems I've solved

Bhavin Kamani

A small command line for a curious visitor.