Image: Composal
Why it mattersA QA team writes a few tests before a release. The agents cover the paths nobody wrote tests for. The bug a user finds first costs more than every test suite ever did.
A team releases a feature and a user finds the bug an hour later. The test suite missed it. Composal is a tool built around that one problem. It launched on Y Combinator Fall 2026. The YC page had 7 votes at the time of writing. The three founders are Pandelis Zembashis, Jonathan Davies and Jan Moehlmann.
The pitch, as the YC launch page states it: "a swarm of agents that tests your product like real users". A team gives Composal a URL, credentials, and sources like the repo, specs and docs. The agents learn the product by using it. They map features. They explore user journeys with different roles. The company calls this "Continual Agent Verification" and places it between CI/CD and production.
What the agents do
The YC page lists four stages. First, discovery: Composal reads the sources the team provides. Second, the agent swarm explores the product and tests key flows in parallel. Third, the output is inspection: for every issue, Composal captures a recording, logs, and the steps to reproduce. Fourth, the reproduction steps go to the team's coding agent, which can then try the fix.
The company's own site has one line that reads "agents build it, ours break it". The product-build agent and the QA agent are the same kind of tool. One writes new code. The other checks that what was released still works. The two sit on opposite sides of the deploy.
Who it is for
The target is any team that already uses an AI coding agent for feature work. The problem Composal names is simple. A coding agent can write a new feature. It cannot promise the paths it did not touch still work. A human QA team can check a handful of paths, not every path. The agent swarm is the attempt at every path. The YC page lists the industry as B2B and the tags as Developer Tools, DevSecOps, Cloud Computing and AI.
What is not known yet
The YC page lists no pricing. It names no customer. It gives no number for how many bugs the agent swarm has found in a real codebase. The CLI uses Anthropic's Claude Opus 5.5. Composal integrates with GitHub, Sentry, PostHog and Notion, which the home page lists. The home page carries the "Backed by Combinator" line. No funding amount is stated. The company's year of incorporation is not given on the launch.
A test for the next six months. Does a Composal team catch more real bugs than a human QA team of the same size? The AI-testing category has a crowded field. Hue, another YC Fall 2026 company, launched a day earlier with a different take on the same ground. Hue simulates an app so an agent can be tested against it. Composal does the reverse: it points agents at a real product. The two can live together. A team can test one agent against Hue's simulator before building Composal's swarm around a real deploy.
The honest limit
The YC launch page is the only primary source for the product on day one. Composal says its agents "autonomously learn the product by using it". How well that generalises across a real app's edge cases is not shown in a launch page. A team that cares should try the swarm on a staging environment with a few known bugs. See which ones the agents find. See which ones they miss.
Source
- Primary source: Composal on Y Combinator.
- Company site: Composal.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
