SurfaceLoom

Appearance

OPEN-SOURCE / AI AGENT TESTING

Test what your
agent actually did.

Approval decisions. Tool calls. Real side effects.
Bring them together in one repeatable test, with evidence you can inspect.

Run the example View on GitHub

Real Chrome / Chromium. No model API key.

CASE EXPLORERIllustrative view
REFERENCE AGENT / 01

A denial you can verify.

Approval decision
Denied
Tool starts
0
Verified local effects
0
PassedComplete ledger

The completed ledger and independent resource probe agree.

REAL BROWSER FIXTUREJSON · HTML · EVIDENCE
THE WORKFLOWApprovalExecutionEffectEvidenceExplore the framework

01 / ASK BETTER QUESTIONS

A plausible answer
is only the beginning.

01

Did “deny” really mean no?

Compare the approval decision with the tool ledger and an independent resource probe.

Test approval flows 02

Did the same thing happen twice?

Check business effects across call IDs. Catch duplicate execution and incomplete observations.

Test tool side effects 03

Can you trust the final result?

Keep unknown execution, missing evidence, and unconfirmed cleanup visible in the report.

Understand the boundaries

02 / THE REFERENCE FIXTURE

Good tests
catch bad behavior.

Four Cases. One deterministic agent. A real, owned browser for each run.

2 passed2 failed

The two failures are intentional fixture faults. They stay red in the report. Expected showcase exit: 1.

Run it. Inspect the evidence.

  1. Denied. Nothing executed.

    Zero starts. Zero effects. Complete evidence.

    Passed

  2. Approved. Executed once.

    One tool execution and one local effect.

    Passed

  3. Denied. Still executed.

    The ledger and resource expose the fault.

    Failed

  4. Denied. Incomplete ledger.

    A missing barrier cannot establish zero.

    Failed

03 / FAMILIAR TOOLS. EXPLICIT BOUNDARIES.

Built around your tools. Grounded in evidence. ​

SurfaceLoom coordinates replaceable backends around AI agent behavior contracts: approval decisions, tool execution, actual side effects, and retained evidence.

For ordinary browser UI tests, Playwright may already cover your needs. SurfaceLoom adds agent behavior assertions and evidence coordination.

See support and limitations →

Browser Playwright
Live approval fixture
Windows UI Automation
Separate C# / UIA fixture
macOS Accessibility
Contract-tested

EXPERIMENTAL No stable API guarantee. Packages are not published to npm. No target-product validation is claimed. The TypeScript-to-native live path is not yet proven.

START SMALL. VERIFY SOMETHING REAL.

Your first test starts here.

Clone the source, run the fixture, and open the report.

Open the quick start

Experimental framework · MIT licensed · No stable API guarantee

SurfaceLoom contributors