Composal

Teams that ship code written by coding agents now produce changes faster than anyone can click through them, and Composal was started in London to close that gap. Given a product's URL and a set of test credentials, it sends a swarm of browser agents through the product's user journeys and edge cases, the way real users would, and reports what breaks.

## Bug reports a coding agent can act on

Each failure comes back with a screen recording, the logs from the run and the steps that reproduce it. Composal then passes the report to a coding agent to fix, and runs the same journey again to confirm the fix holds. Optional connections to GitHub, Sentry, PostHog and Notion give the agents context on the code, the errors users hit, real usage and the team's own specifications.

According to its documentation, tests can run from a command-line tool, against preview deployments opened by pull requests on GitHub or GitLab, and through an MCP server so other agents can call it. Private staging environments are reached over Tailscale or AWS Client VPN, and test logins are held as managed credentials rather than pasted into prompts.

## The team behind it

Composal is part of Y Combinator's Fall 2026 batch and lists three people. Pandelis Zembashis, the chief executive, was chief technology officer of Cosine and before that head of engineering at 11x. Jonathan Davies, the chief technology officer, co-founded the business consultancy mmtm and later Papaya. Jan Moehlmann, the chief revenue officer, held commercial roles at 11x, Encord and Tractable.

## Key facts

**Maker:** Composal, London, United Kingdom **Launched:** 30 October 2026 on Y Combinator's launch board (Fall 2026 batch) **Pricing:** not published; sign-up or a booked call **Integrations:** GitHub, GitLab, Sentry, PostHog, Notion, MCP, Tailscale, AWS Client VPN **Website:** composal.com

Not yet independently verified: Composal launched on 30 October 2026 and names no customers; everything here comes from its own site, documentation and YC launch page, read on 30 October 2026. We have not seen an independent test of how well its agents find real bugs. We will update this when it can be confirmed, and remove this note.