Momentic Introduces Agents That Catch Bugs in Web, iOS, Android Apps

Wei-Wei and Jeff, co-founders of Momentic, a Y Combinator W24 startup, have launched a suite of QA agents designed to autonomously plan, explore, and identify bugs in web, iOS, Android, and upcoming desktop applications. The founders, who previously worked on development tooling and infrastructure at companies including Robinhood, Retool, WeWork, and Qualtrics, developed the platform to address the inefficiencies of traditional testing frameworks like Playwright, Selenium, and Maestro. They argue that these existing tools require hardcoding implementation details and are misaligned with modern agentic software development life cycles, particularly as the adoption of coding agents increases the change failure rate.

Momentic’s core product, named Mo, is a custom agent harness built on three years of development and millions of test traces from real-world applications. The system utilizes a self-organizing agent graph that builds a directed acyclic graph during exploration, automatically forking large product areas into parallel sub-agents for concurrent execution. It employs specialized agents for domain-specific tasks, including planner agents that reason over screenshots and video, and locator agents that inspect page structure to resolve plain-English interaction targets. Additionally, the infrastructure supports cooperative resource leases, allowing multiple agents to coordinate when using or modifying overlapping resources, and includes a tunnel for exposing private sites and local apps to hosted agents.

In a benchmark comparing Mo against other coding agents on four open-source applications—Gitea, Metabase, Twenty, and Saleor—the startup reported that Mo achieved higher recall rates. The evaluation involved 1,396 human-labeled bugs, with each agent given the same brief and a fresh sandbox without source code access. In a time-boxed 60-minute run, Mo identified 269 bugs with an 84% precision rate and a 19.3% recall rate. By comparison, Codex found 86 bugs with 86% precision, Claude Code found 59 bugs with 80% precision, and Devin found 27 bugs with 84% precision. The cost per bug for Mo was $2.67, which is higher than Codex’s $0.80 but lower than Claude Code’s $4.09 and Devin’s $3.89.

The founders note that Mo’s performance scales with time, increasing its bug count by 52% between 30 and 60 minutes, whereas Codex saw only a 16% increase over the same period. While Mo prioritizes recall to find the most real bugs, they acknowledge that Codex is slightly more precise. The company has recently shipped PR testing capabilities for Vercel and Cloudflare deploy previews. New users are currently offered $250 in Mo spend, while paid plans provide higher infrastructure capacity and included usage options. The startup has published its results, write-up, dataset, and harness code for public review.