Kane CLI: No. 1 Software Testing Agent [PRISM Q4 2026 Report]

What PRISM Measures, and Why It Matters for Agentic QA

PRISM asks whether a testing agent behaved correctly, not just whether a test finished. Automated tests read the DOM, but people read the screen. PRISM measures the gap between the two.

Every agent runs the same 50 scenarios, five in each of ten categories. They range from hydration and timing races to canvas and WebGL, shadow DOM, layout shifts, generative UI, and long multi-step flows. Each scenario is scored on three dimensions:

• Reliability (R): does the agent get a deterministic, repeatable result?
• Perception (S): did it actually see the state a person would see? This is a pass-or-fail gate.
• Intent (I): did it do what the task asked, and only that?

The RPS-Index is the product of the three averages (R × S × I). A weak score on any one dimension pulls the whole index down, so an agent can't hide poor perception behind high reliability.

PRISM also counts three failure types that matter most in Agentic QA:

• False Heals: the agent reports a pass while the sandbox confirms the flow is broken. In other words, a green test over a broken checkout.
• Intent violations: the agent clicked something it shouldn't have, or took an action nobody asked for.
• Integrity violations: the agent gamed the measurement, arriving at an answer by a route a real user could not take. These void the scenario outright.

The scenario specs and scoring code are public on GitHub. The sandbox is not, so no vendor can tune its agent against the exact pages on which it is measured. Every vendor can also publish a reply of up to 100 words next to its row. Full details are in the methodology.

The PRISM Q4 2026 Leaderboard

Seven agentic testing tools were ranked this quarter. Kane CLI finished first, with Momentic a close second.

Source: PRISM Q4 2026 results.

No entry recorded an integrity violation. "Operator baseline" means PRISM's maintainers ran that tool themselves.

The gap between plain Playwright (0.46) and the top agents (0.76+) is the clearest finding on the board. A scripted framework reads the DOM. An agentic testing tool that perceives the page the way a user does catches what a script misses.

Where Kane CLI Led

Source: PRISM Q4 2026, Kane CLI 0.8.17 results page · 10 categories × 5 scenarios

Kane CLI was perfect in two categories: hydration and timing races, and affective and cognitive load. Timing races are where most flaky e2e tests come from, so this is the result we're proudest of. It beat its own overall score in five of ten categories, and recorded zero integrity violations across all 50 scenarios.

What We're Still Fixing

A No. 1 rank is not a perfect score, and PRISM makes that easy to see. Three numbers on our row are now our Q1 2027 targets.

• False Heals: 9 of 50. Momentic recorded 5 and Shiplight CLI 3. A false heal is the most expensive failure in AI testing because it hides a real bug behind a green check. Cutting this number is our top priority.
• Peak complexity and multi-step logic: 0.40. This is our weakest category. Long flows with branching logic are where intent drifts, and it produced 2 of our 3 intent violations.
• The margin is small. We lead Momentic by 0.0027 on the RPS-Index. That is a lead, not a gap, and next quarter's board is open to everyone.

This is why we run independent teardowns: free licenses to anyone who wants to break Kane CLI and publish what they find. Much of the improvement that put Kane CLI on top came from teams telling us exactly where it broke.

Why This Matters When Choosing Agentic QA Testing Tools

Coding agents like Claude Code, Cursor, Codex, and Gemini now ship features in hours. Every one of those changes still has to be verified before it reaches a customer. Verification is the new bottleneck in software delivery, and agentic testing is the solution.

Agentic testing means an AI agent plans, runs, and fixes end-to-end (e2e) tests based on a goal stated in plain language, rather than having a human deal with brittle selectors. The problem is that an agent that "heals" too eagerly ends up treating real bugs as green checks. That is why a benchmark that evaluates perception and intent and tracks false heals is more useful than a simple demonstration.

When you compare agentic QA testing tools, ask four questions PRISM helps answer:

• Can it see what the users see? With regard to canvas, the shadow DOM, and content that loads later, DOM-only tools lose the ability to see them.
• Does it stick to the task? An agent that confidently clicks the wrong control is worse than one that fails honestly.
• How frequently does it give false results? Instead, look at the number of false heals, not just the pass rate.
• Will you keep the tests? Ensure that the results are exported to your own code, for example Playwright, so that you are not effectively renting your test suite.

How Teams Use Kane CLI

Kane CLI is the AI testing agent for developers and SDETs. It runs in the terminal and plugs into the coding agents your team already uses, including Claude Code, Codex CLI, Cursor, and Gemini CLI. The agent that wrote a feature hands it off to an independent agent for review.

• Intent-based browser control. Describe the flow in plain English. No selectors to maintain.
• Runs that finish. It adapts via UI changes for up to 50 steps per flow until the entire journey is verified.
• Playwright export. Turn any plain-English flow into native Playwright code with one command.
• Built for CI. Pass/fail results with step traces and screenshots on every pull request.

Kane CLI is part of TestMu AI's Agentic Quality Engineering platform. KaneAI covers no-code test authoring for testers and product managers. Agent Testing covers chat, voice, and IVR agents, and HyperExecute runs it all at scale on real browsers and devices.

Download Kane CLI and run it against your own app. Then tell us where it breaks.