Warp Factories: Agent Pipelines You Version Control
Warp opened a closed beta of Factories, version-controlled pipelines that move a ticket through triage, spec, implementation, review and verification. Its CEO puts end-to-end automation at 30 to 35% of weekly tasks internally, which is the most useful number in the launch.
Warp opened a closed beta of Factories on 18 August: cloud infrastructure for running a "software factory", a repeatable pipeline where coding agents take a ticket through triage, specification, implementation, review and verification. Each factory is defined in version-controlled configuration, so the pipeline gets reviewed and changed like code rather than clicked together in a UI.
The launch is worth reading closely for one number Warp's CEO gave TechCrunch: "We automate like 30% of our tasks, 30 to 35% on a weekly basis." That is a company shipping factory infrastructure telling you what a factory currently handles on its own. It is a useful ceiling to hold in mind, and it is not the number most agent marketing implies.
What a factory actually is
Strip out the product and the pattern is five stages, each with a defined input and a defined output:
Triage. A ticket arrives and gets classified. Is it a bug, a feature, a chore? Which part of the codebase does it touch? Is it even actionable?
Specification. The agent writes what it intends to do before it does anything. This is the human checkpoint that matters most.
Implementation. Code gets written.
Review. A second pass, usually a different agent or a different prompt, reads the diff against the spec.
Verification. Tests, builds, whatever proves the change does what the spec said.
Work enters through the tools teams already use, ticketing and messaging systems, and results flow back to wherever the request came from. Zach Lloyd's framing of the hard part is honest: "running your agents in the cloud and steering those agents as they run is actually a huge infrastructure undertaking." The agent loop is the easy bit. Everything around it is the product.
The stage that decides whether it works
Look at the five stages again and notice which one is not about generating code.
Triage, specification, review and verification are all about deciding, describing and checking. Only stage three writes anything. That ratio is the whole point, and it is why "the agent writes code fast" stopped being the interesting property somewhere around last year. The constraint on shipping moved to review capacity, which is the same thing Linear's pull request data showed from the other direction.
A factory is really a machine for making agent output reviewable at volume. If your review stage is a human reading a 900-line diff with no spec to compare it against, adding pipeline infrastructure in front of that will make things worse, faster.
Building the same loop without the product
You do not need cloud factory infrastructure to run this pattern on a two-person team. You need four artifacts, and you can have all of them by Friday.
Artifact | What it does | Where it lives |
|---|---|---|
A rules file | Tells every agent your conventions once instead of per prompt |
|
A spec template | Forces the agent to state intent before acting | A markdown file the agent fills in |
A review prompt | A second pass that reads the diff against the spec | A saved prompt, not improvised |
A verification command | One command that proves it works | Your existing test or build script |
That is a factory. The version-controlled config that Warp sells is the same idea at a scale where "a saved prompt" stops being enough, and where you need the pipeline itself to be reviewable because ten people depend on it.
If you want the individual pieces: how to write an AGENTS.md file agents actually follow covers the rules file, reviewing an agent plan before it runs covers the spec checkpoint, and running an AI coding agent in CI covers the verification stage. The AI coding tools guide is the map if you are still choosing the agent itself.
What this does not solve
Worth stating plainly, because the word "factory" carries an implication the product does not claim. TechCrunch notes Factories is not designed to replace engineers; it facilitates collaboration with agents.
The 30 to 35% figure is the honest read. Roughly a third of tasks go through end to end. The rest need a person somewhere in the loop, usually at specification or review, and those are exactly the two stages that do not get faster when the model gets better at writing code. Lloyd expects the number to rise as models, context and harnesses improve. Fine. Plan for a third today.
The other thing a pipeline does not fix is task sizing. Feed a factory a ticket that says "improve onboarding" and you get five stages of confident nonsense. Feed it a ticket with a testable outcome and it works. That is a prompt discipline problem, not an infrastructure one, and it is the cheapest place to improve your results this week. Running multiple coding agents in parallel runs into the same wall from a different angle.
FAQ
What is Warp Factories? Cloud infrastructure for running version-controlled agent pipelines that move a ticket through triage, specification, implementation, review and verification, with human checkpoints. It opened in closed beta on 18 August 2026, access by request.
Which models does it run on? It is model-agnostic, working with harnesses including Codex and Claude Code, with ticketing and messaging integrations for Linear, Jira, Slack and Teams.
Does a software factory replace developers? No, and Warp does not claim it does. Its CEO puts current end-to-end automation at 30 to 35% of weekly tasks internally, with people still required at specification and review for the rest.
Can a small team do this without buying anything? Yes. A rules file, a spec template, a saved review prompt and one verification command give you the same loop. The paid infrastructure earns its cost when the pipeline itself needs to be reviewable by a team.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
