Warp Factories turns tickets into PRs — keep the harness open

Warp launched Warp Factories into closed beta on August 18 — infrastructure for running fleets of cloud coding agents that take a backlog ticket and hand back a reviewed pull request. The interesting part is not the demo. It is that Warp built the thing to run someone else's agent, and aimed it at companies too small to build their own.

What actually happened

Per Warp's announcement, a factory orchestrates work through four phases, with a "foreman" agent routing between them by complexity:

• Triage — decide whether to implement, spec, or defer the item
• Spec — pull a human in when the change is ambiguous
• Implement — write the code
• Review — evaluate the result, with computer-use verification so agents test their own work

The platform is multi-model and multi-harness on purpose: TechCrunch reports it works with Codex and Claude Code alike, and Warp lists open-weight models alongside them. It ships cost, quality, and ROI dashboards, custom scoring, and observer agents that tune factory configuration over time.

Two numbers worth holding. Warp says factories already handle about 30% of its own internal tasks — a vendor-reported figure about a vendor's own codebase, so treat it as a direction, not a benchmark. And qualified organizations joining the closed beta get $10,000 of factory use free. Standard pricing is not public yet.

Why it matters for your business

We run coding agents on client work, so here is the operator read.

Model-agnostic is the feature. Most agent platforms want to be the model, the harness, and the pipeline. This one sells the pipeline and lets you swap the rest. That matters when your provider raises prices, degrades a tier, or ships a model that regresses on your codebase — and all three happened this year. Portability is not idealism. It is the thing that keeps a repricing from becoming a rewrite.

Ticket-to-PR is a governance problem before it is a throughput problem. The phase you should stare at is triage, because "defer" is the only output that costs nothing and the only one an agent has no incentive to pick. Decide in advance which ticket classes are allowed to reach implement without a human on spec. Anything touching payments, auth, or migrations should not be on that list.

Free credits are a measurement window, not a discount. $10k of factory use is enough to answer one question honestly: on your repo, what fraction of merged PRs from the factory needed rework? Log it. A factory that produces plausible PRs your seniors rewrite is a cost transfer, not automation.

Dashboards measure the pipeline, not the outcome. Cost per task is easy. Cost per shipped, unreverted change is the number that decides whether this works. Instrument that yourself — no vendor dashboard will do it for you, because the answer is not always flattering.

Sources: Warp — Introducing Warp Factories, TechCrunch.