Momentic Uses AI Agents to Explore Apps and Find Bugs

Introduction

Software teams are producing code faster, but faster code generation does not automatically produce more reliable software. Momentic, a YC W24 startup, is trying to move large language models beyond code generation and into software testing. Its QA agents are designed to operate Web, iOS, and Android applications, plan actions, explore interfaces and user journeys, and report possible bugs.

Key points

• Testing by using the product: Rather than only generating assertions or test scripts, Momentic’s agents are intended to enter an application, interact with controls, complete flows, and choose subsequent actions based on what they observe.
• Less dependence on brittle selectors: The team argues that tools such as Playwright, Selenium, and Maestro often require teams to encode XPath, text, or aria attributes directly into tests. When an interface changes, those scripts can require repeated maintenance. Momentic wants agents to rely more on task goals and visual or semantic understanding.
• Coverage for user journeys: The company presents its system as a way to inspect controls and end-to-end flows. The supplied example describes a consumer fintech application where a login-routing change skipped a new-device verification prompt, silently bypassing a security checkpoint.
• A response to coding-agent adoption: The founders previously worked on developer tooling and infrastructure at companies including Robinhood, Retool, WeWork, and Qualtrics. Their argument is that faster changes can expose weaknesses in testing if validation does not scale with development speed.

Why it matters

Conventional test automation is valuable because it is deterministic and repeatable. Its trade-off is that engineers must often understand an application’s structure in advance and encode those details into scripts. An agent-based approach starts from a task—such as logging in or completing a user flow—and tries to find a workable path dynamically. In principle, this could help uncover combinations of actions that were never explicitly written as test cases.

The approach also creates a different set of evaluation challenges. It is not enough for an agent to click through a flow once. Teams need to know whether important paths were covered, whether security issues are missed, and whether reported failures can be reproduced with useful evidence. In its launch material, Momentic cites a DeepSWE v1.1 comparison in which the stated pass rate was 66.2% when the model was prevented from writing tests, versus 65.3% when it was allowed to write them. The figure is presented by the company as support for its argument that generating tests does not automatically improve validation; independent evaluation would still be needed.

Momentic is therefore not necessarily replacing Playwright or Selenium. Its more specific ambition is to delegate the labor-intensive parts of exploratory testing, test maintenance, and bug discovery to an agent. If such systems can provide reproducible steps, convincing evidence, and a low false-positive rate, they could add a layer of active exploration alongside existing CI pipelines. For now, the public material consists mainly of the company’s launch post and demonstration, so its reliability and depth of coverage in production environments remain open questions.

Hacker News