For AI agents: the complete documentation index is available at https://a3s-lab.github.io/Test/v1.0.1/en/llms.txt, the full documentation bundle is available at https://a3s-lab.github.io/Test/v1.0.1/en/llms-full.txt, and this page is available as Markdown at https://a3s-lab.github.io/Test/v1.0.1/en/index.md.
A3S Test
  • English
  • v1.0.1

      • Ground the agent in the real page.Keep proof of every change.

      Select an interface problem, carry the current revision, component, and source hints to the agent, then verify the result against the changed page. A3S Test turns that feedback loop into an executable, inspectable test.

      Run your first test
      A3S ShopInteractive demo

      Review order

      Interactive demo
      1. 1Cart
      2. 2Review
      3. 3Payment
      4. 4Complete

      Delivery details

      Alex Chen · +1 555 0100

      1001 Interface Way, San Francisco, CA 94107

      Order items

      Quantity 1
      A3S autonomous testing suiteTeam · annual license
      $1,298.00

      Page context

      Reading the real page context…

      1. 1Read page
      2. 2Bind target
      3. 3Describe issue
      4. 4Build batch
      5. 5Save in tab

      Read the current render

      Semantics, state, geometry, and page revision come from the running browser.

      Bind the owning source

      Components, locators, regions, and source candidates travel with one finding.

      Accept only fresh proof

      Verify against the changed page, then preserve the stable path as ACL.

      Connect intent, interface, and evidence into one replayable path

      The complete flow starts with the PRD, design reference, and running page together. The coding agent understands and plans; A3S Test constrains each browser action, retains evidence, and turns only proven paths into deterministic regressions.

      checkout-ui · workflow walkthrougha3s.test.workflow/1Select a phase manually

      Place the PRD, design reference, and running page in one context

      Requirements define the outcome, the design reference constrains visual intent, and the browser provides current facts. Each input keeps its own boundary.

      INPUT CONTRACT

      Three inputs

      • checkout-requirements.mdbusiness rules · observable success
      • design-reference.pngvisual target · layout intent
      • localhost:3000/checkoutrunning UI · current revision
      Build the task contract

      First define what the page can prove

      Goal
      Reach confirmation after submit
      Success
      “Order confirmed” heading is visible
      Scope
      /checkout · same-origin page
      Authority
      Read evidence; no source-edit grant

      Task contract

      READY
      goalobservablereferenceboundedtargetrunning pageauthorityunchanged

      Every run keeps the smallest reviewable evidence set

      SCREENSHOTfinal visual stateACCESSIBILITYsemantics and operabilityCONSOLEbrowser log streamPAGE ERRORSuncaught runtime failuresHAR · TRACE · VIDEOcaptured only when needed

      Failures are classified as product / specification / infrastructure. HAR, trace, and video stay opt-in. Every session ends with finish or abort and cleans up the resources it owns.

      Install the execution core with one command

      The CLI and Agent Skill run sessions and ACL. Add Test Kit only when selecting, drawing, capturing, or sketching inside the page helps.

      Add Web Test Kit when needed
      CLI + Agent Skill · v1.0.1
      curl -fsSL https://github.com/A3S-Lab/Test/releases/latest/download/install.sh | sh -s -- --version v1.0.1

      The installer verifies the release archive SHA-256 and keeps the CLI and Skill on one version.

      Same tasks. Compare completion, safety, and cost.

      This comparison isolates the execution layer: both candidates use one deterministic solver and the same browser foundation, with official MiniWoB reward as the only success verdict.

      Locked MiniWoB++ protocol · 9 tasks × 3 repetitions · Apple M2 Pro · 332 × 214run 20260821T171716Z
      UI execution-layer comparison between A3S Test and direct agent-browser
      MeasureA3S Test 1.0.0Direct agent-browser 0.26.0
      Task success100%27 / 27 main runs88.9%24 / 27 main runs
      Stale-reference protection3 / 3old refs rejected · 0 page mutations0 / 3old refs rejected · 3 page mutations
      Built-in run evidence27 / 27runs with built-in evidence files0 / 27runs with built-in evidence files
      Paired common-success median7.48 s+10.6% versus the direct client6.74 shost-local baseline

      This is single-host development evidence, not a model-reasoning or universal-performance claim. Drag and wheel probes timed out for both candidates and are excluded from the main success rate.

      Read the full protocol, raw data, and reproduction steps

      Give the agent more than “something is wrong here”

      Test Kit combines human intent with bounded context the browser already computed. The agent can see the target, its likely owner, and the exact page revision that produced the evidence.

      Inspect Page Context, authority, and fail-closed rules
      Example handoff packet
      protocol    a3s.test.repair/1
page        /checkout · revision 18
selection   button “Submit order”
source      TestKitExperience.tsx
intent      Increase text contrast
context     role · locator · viewport geometry
authority   human submitted

      Page-derived content remains untrusted evidence. Selecting, capturing, or saving a draft never grants source-edit authority.

      Run one real page in three steps

      Keep the application running as usual, then open a session around one explicit goal. Execute one typed action at a time from the latest observation.

      1. Run your page

        Use the project command you already have. A3S Test does not need a custom build step.

        npm run dev

      2. Start a session with a visible outcome

        The goal and success condition must be observable from the page.

        a3s-test agent start http://127.0.0.1:3000 \
  --session ui \
  --goal "Reproduce the submit failure" \
  --success "The confirmation heading is visible"

      3. Observe before acting

        Use fresh refs, verify the result, then distill the stable path into ACL.

        a3s-test agent observe --session ui --interactive --json
      Open the complete quick start

      Answers that should shape the decision

      Clear boundaries matter more than putting every capability on one page.

      Start with one real page

      Prove one session with an observable success condition. Add Test Kit only when richer page context helps.

      Run your first testAdd Test Kit