Momentic Mo QA Agent Catches 84% of Bugs Autonomously

YC Winter 2024 startup Momentic has officially released Mo, an autonomous QA agent that plans, explores, and bug-bashes applications across web, iOS, Android, and desktop platforms. The launch includes QA Bench, a new benchmark showing Mo caught 269 of 1,396 human-labeled bugs at 84% precision, significantly outperforming competing coding agents including Codex, Claude Code, and Devin.
QA Bench Benchmark Results
Momentic tested Mo against three competing agents on four open-source applications: Gitea, Metabase, Twenty, and Saleor. The benchmark covered 20 product areas and 1,396 manually labeled bugs collected by Momentic's team over three years of testing traces.
Mo matched 269 bugs with 19.3% recall and 84% precision at $2.67 per confirmed bug. Codex with GPT-6.1 Sol and Playwright MCP finished second with 86 matches at 86% precision and $0.80 per bug. Claude Code using Opus 5.5 found 59 bugs at $4.09 each, while Devin identified 27 bugs at $3.89 per match. Mo's match count increased 52% when runtime extended from 30 to 60 minutes, compared to just 16% improvement for Codex over the same period.
Release Date and Availability
Officially launched on November 2, 2026, Mo is now serving enterprise customers including Notion, Runway, Webflow, Quora, and Retool. The company published both the agent and the QA Bench harness alongside a Launch HN post detailing the architecture and benchmark methodology.
Self-Organizing Agent Architecture
Mo's technical foundation uses self-organizing agent graphs that fork into parallel sub-agents with cooperative resource leases, preventing test collisions when multiple threads target the same fixtures. Co-founders Wei-Wei Wu and Jeff An, who previously built developer tooling at Robinhood, Retool, WeWork, and Qualtrics, described the system as capable of autonomously exploring user flows through plain-English locator agents operating across a directed acyclic graph of forked browsers and devices.
The architecture represents what one Launch HN commenter characterized as a non-trivial orchestration challenge. Others noted the familiar tension in agent tooling, where custom harnesses risk obsolescence as frontier models advance. Mo's performance gap over Codex widened with additional runtime, suggesting the orchestration layer captures value beyond raw model capability.
Market Context and Competition
The Mo launch arrives amid a crowded field of 125 coding-tool announcements tracked by AI Weekly over the past 90 days. That same period saw Goldman Sachs remove Claude Code from trading desks, highlighting enterprise caution around autonomous coding agents in production environments.
The 1,396-bug ground truth used in QA Bench comes from Momentic's internal corpus rather than an independent third-party dataset. The benchmark focuses specifically on functional bugs discoverable through UI exploration, distinct from unit-test generation or static analysis tasks where other coding agents concentrate.
What This Means
Momentic's launch establishes a new benchmark specifically targeting autonomous QA exploration, an area where general-purpose coding agents have shown limited effectiveness. The 84% precision rate and scalable cost structure position Mo as a practical tool for continuous testing pipelines, though the proprietary ground-truth dataset and customer overlap between Momentic and benchmark subjects raise questions about generalization to new codebases. The agent's performance improvement with extended runtime suggests diminishing returns may favor hybrid workflows pairing Mo's exploration with human triage over fully autonomous testing at scale.
on Emergent today





