AI product
Warp Factories is a platform for defining and running cloud software factories: fleets of agents that automate software-development workflows from task intake through review and a mergeable pull request. Factories are specified as code in a factory.yaml file, including repositories, agents, models or harnesses, runners, credentials, MCP servers, cloud access, integrations, triggers, and approval gates. Agents run in isolated execution environments, while Warp centralizes their transcripts, tool calls, outputs, and traces for evaluation. Evals and LLM-as-judge rubrics can score factors such as redundant tests, verbosity, and efficiency; benchmark runs compare configurations, and self-improvement agents identify recurring failures and open pull requests that modify factory skills or configuration. The platform supports models and agent harnesses such as Claude Code and Codex, integrations with work and source-control systems, and deployment in Warp's cloud or a customer's own VPC.
2 uses taken from transcripts — each links to the moment in the video.
Platform for building software factories: runs agents in isolated Docker sandboxes, centralizes agent traces, hosts scoring and self-improvement agents, and presents proposed self-improvement PRs in a review inbox.
The product from Warp demonstrated in the talk: a cloud software factory platform where the factory is defined in code (agent definitions, workflows, automations, benchmarks, skills, webhooks stored in a repo), work flows in from tools like Slack, Linear and GitHub, all agent runs are recorded and scored by LLM-as-judge agents, failures feed a self-improvement loop that proposes fixes, and you can benchmark past tasks against different models to build model routing strategies.
2 in the library.