Home›Companies›Composal
Composal

Always-on agents that keep working after engineers log off

Composal gives engineers always-on agents that work in the cloud, each with its own computer, Slack handle and memory. An engineer gives one a standing order, like “watch Sentry, fix new errors, get the fix reviewed and ship it,” and it keeps going after the laptop closes. Our QA agents test every pull request by using the app like a real user, and have found over 1,500 bugs in customers’ apps since mid-October.
Active Founders
Co-founder and CEO of Composal (YC F26). Engineers now spend their days managing threads of AI agents. Our always-on agents take a standing order and own the work, so engineers only hear back when it’s done. Previously CTO at Cosine (YC W23), where I built the harness and sandboxes behind an AI software engineer that topped SWE-bench in August 2024.
Jonathan Davies
Jonathan Davies
Founder/CTO
Co-Founder and CTO of Composal. Previously Senior Software Engineer at PapayaDash. Before that built software for Telefonica, Mitie, and Vodafone
Founder at composal.ai (YC F26). Previously built commercial functions at 11x, Encord, Tractable
Company Launches
Composal: A swarm of agents that tests your product like real users
See original launch post

Composal sends a swarm of agents through your product to check that users can still do what they came to do. When something breaks, it captures a recording, logs and steps to reproduce, so your coding agent can make the fix and Composal can verify it.

We're Pandelis, Jonathan and Jan, the founders of Composal (YC F26). We're building the verification scaffolding for autonomous software factories: the part of the loop that is still reliant on humans today, checking whether what the agents built actually works.

https://youtu.be/MmnVcWLlnAQ

The Problem

Coding agents have made building cheap. Checking has not kept up. Teams tell us about features that pass CI and get approved by a review bot, then fail the first time a real person tries them. So someone still clicks through every flow by hand before a release, or skips it and hears about the bug from a customer.

End-to-end test suites were supposed to solve this. In practice they are tedious to write, they break every time the UI changes, and they are the first thing to rot. As agents raise the pace of shipping, that gap widens every week.

How Composal works

Give Composal an environment to test and the credentials it needs. Our agents learn your product by using it, mapping features and exploring user journeys in parallel with different roles and goals.

For example, testing an invitation flow means following what happens to the person receiving the invite: can they join the right workspace and reach the features they need? A successful API response only answers part of that question.

When an agent finds a problem, Composal captures the evidence. Your coding agent uses it to investigate and make a fix, then runs the scenario again to confirm the result. What Composal learns becomes product knowledge and repeatable scenarios that keep running as your product changes.

We call this Continual Agent Verification: checking that users can still accomplish their goals as the product evolves.

uploaded image

CI checks code, CD ships it, Continual Agent Verification checks that user goals still work as the product changes.

Why we're working on this

Pandelis spent the last few years building coding agents as CTO at Cosine (YC W23). That team was no different from the teams we talk to now: ship a feature, test what changed, move on, then onboard a customer and find that a flow nobody had touched in weeks was broken. The agents kept getting faster. The checking stayed manual. Composal is the tool we wished we had.

Who we are

  • Pandelis Zembashis: previously CTO at Cosine, where he built one of the first CLI agent harnesses, sandboxed execution from scratch, and large-scale repository mirroring for model training.
  • Jonathan Davies: built mission-critical software for Telefónica, Vodafone and Mitie.
  • Jan Moehlmann: built commercial functions at 11x, Encord (YC W21) and Tractable.

We worked closely together at previous startups in London, and we keep hearing the same question from customers: how do you know an agent's work is ready for your users?

Where this goes

Our ambition is a self-healing software factory: you set the direction and agents carry the work from idea through building, testing and shipping, then keep improving the product. Verification is the first step, because nothing else in that loop can be trusted without it.

Build, Verify, Ship, Improve: a fault is caught and healed without stopping the loop.

Try it on a feature you're about to ship

If your team uses coding agents and still checks their work by hand, we'd like to test an upcoming feature with you.

Start at composal.com to run your first discovery, or book a call with us directly and we would be happy to setup your first batch of tests for you.

— Pandelis, Jonathan and Jan

YC Photos
Composal
Founded:2026
Batch:Fall 2026
Team Size:3
Status:
Active
Location:London, United Kingdom
Primary Partner:Pete Koomen