Momentic launches AI QA agents it says caught over 117,000 bugs in 70 million test runs

What launched

Momentic, a Y Combinator Winter 2024 startup, has launched on Hacker News with a product it describes as an AI QA engineer, named Mo, that tests applications end to end from a prompt describing a user flow. Instead of recording scripted UI tests built on fragile selectors, engineers describe in plain English what should happen, and agents drive the app to check it.

The launch post, dated November 2, 2026, argues that coding agents have made building software dramatically faster, and that the bottleneck has moved to knowing whether a change actually works. Momentic's position is that verification should happen inside the same loop, with agents gathering evidence directly from the application.

How the agents work

Tests are defined as plain-English steps and assertions stored in YAML files, which live in the repository so engineers and coding agents can review and version them. Around that core, the platform layers several supporting systems.

A knowledge base holds product terminology, business rules, and expected behavior; AI-assisted steps consult the relevant entries while a test runs. An app graph maps the screens and user flows the agents encounter during executions.

When the interface changes, the agents try to relocate controls mid-run. If the change is permanent, the system proposes an edit to the test file for a person to review, apply, and rerun. Momentic reports 8,932,104 of these auto-healing actions to date.

For failures, the tool shows which step broke, along with screenshots and traces, and an AI triage pass classifies each failure as an application bug, test flakiness, or an environment problem.

Integration spans the development workflow. Coding agents can connect through MCP to author a test for a changed flow, try it against a live session, and save the resulting YAML next to the source. A CLI runs the same tests inside CI, and its exit code can act as a required check before a merge or deploy.

The published numbers

The company's site displays running totals, which are self-reported rather than independently audited:

• 77,813 tests created
• 70,607,819 test runs executed
• 339,166 changes verified
• 80,290 pull requests verified
• 8,932,104 auto-heals
• 117,010 bugs caught

Named customers appear in testimonials on the site, including Notion, Webflow, Runway, Quora, and Coframe. A Quora case study cited by Momentic reports daily test execution dropping from seven hours to 30 minutes, more than 500 manual test cases replaced, and all critical tests created within a month. Runway's engineering manager describes an automated evaluation across 110 prompts that finished in 14 minutes, a task he estimates would have taken well over an hour manually. Coframe says it runs end-to-end QA on every pull request, parallelized 6.3 times over.

The bugs it claims to have caught

The site lists anonymized examples along with the stage where each was detected. They include a login routing change at a consumer fintech that silently skipped the new-device two-factor verification step, caught on a feature branch; a payment-form refactor at a food-delivery marketplace that stopped pre-filling saved cards at checkout, flagged eight minutes before a deploy; a dependency update at a large rideshare app that left the driver ETA stuck on "calculating" after booking; and twelve legacy API issues found in a single day ahead of a major release.

Why it matters

End-to-end testing has long been the least loved part of the toolchain: expensive to write, quick to rot, and flaky enough that teams stop trusting CI. AI coding agents make that pressure worse, since more code ships faster with less human review per change. Momentic is betting that verification itself becomes an agentic task: tests written in natural language, healed automatically as the UI shifts, and consumed by the same coding agents that produced the change.

The published totals, if accurate, point to real adoption rather than a demo, though they are the company's own figures paired with customer quotes from its marketing site. The open question is reliability at scale, and whether AI-driven triage and healing stay dependable enough that teams will gate merges and deploys on an agent's exit code.