Warp launches Factories for AI dev teams

Warp launched Factories on August 18, infrastructure for running what it calls “AI software factories,” agent loops built around the standard stages of development: triage, specification, implementation, review, and verification. The system hosts agents in the cloud, lets a developer steer them and move work to a local machine when needed, and plugs into Linear, Jira, Slack, and Teams so agents can pick up tickets and report back where a team already works. It also tracks performance and token spend per task and runs self-improvement loops to tune how the factory operates over time. Factories work with multiple coding models, including Codex and Claude Code, rather than locking teams into Warp’s own models. Access is available on request; Warp hasn’t published pricing yet.

Warp CEO Zach Lloyd said building this kind of setup from scratch, running agents in the cloud, steering them, moving work locally, setting up memory and evals, is “a huge infrastructure undertaking to do this right,” and pitched Factories at smaller companies that can’t build it themselves. He said Warp’s own agents currently automate 30-35% of tasks in a given week, with that share expected to grow as models and context windows improve. It’s a bigger bet than Warp’s standalone Agent CLI shipped two weeks earlier, moving from a single terminal tool to the infrastructure meant to run a whole team’s agent workflow.

What it means for you

The 30-35% figure from Warp’s own usage is the honest baseline to compare against if a vendor pitches you a much higher automation number without showing their work. Factories is a bet that most teams don’t want to hand-build agent orchestration, similar to the bet behind Cursor’s new Origin code host, just aimed at the pipeline around the agent instead of the repo it writes to. If you’re already running Codex or Claude Code informally across a team, evaluate whether the ticketing and spend-tracking layer here saves more time than it costs to adopt, since that operational layer, not raw model quality, is usually what breaks first at scale.