Warp

Filed by the Vibe Code Weekly desk · Updated Nov 12, 2026

The story

Warp, which also offers a terminal download, now sells a control layer for teams running many coding agents at once. Each factory is written as a config file, accepts the agent and model a team prefers, and grades its own runs so the setup can be tuned over time rather than left to guesswork.

Warp is aiming below the assistants, at the plumbing that sits underneath those assistants. The company describes this layer as infrastructure for software factories: groups of agents that pick up tasks, work in real repositories and hand back pull requests. Warp says it is built as a platform rather than a vertical product or an AI teammate, with an API, CLI, SDK and MCP access. Claude Code, Codex, Cursor and Warp's own agent are listed as compatible, alongside frontier and open-weight models. That makes Warp less a rival to those agents than a way to coordinate them.

The factory itself lives in a factory.yaml file. In Warp's sample, a web-server factory names two repositories under one owner, a deploy token secret, a Sentry MCP connection, an AWS IAM role and Slack and Linear integrations. The alias field sets the handle of the Foreman, the coordinating agent teammates can @-mention in Slack or Linear to hand work to the factory. Roles inherit a default model, which can be set to auto, and a runner, and individual agents can swap either. Because the definition is code, changes to cost or performance settings can be versioned and compared.

The other half of the pitch is feedback. Warp scores each run against evals built from a team's own work, benchmarks models side by side and runs loops in which observer agents propose edits to the factory config as pull requests. Warp links to a case study on lowering cost per PR from $80 to $30 by testing models with its benchmarks. Customer stories describe Rectangle Health building Rex, an AI teammate on Oz that now writes 54% of its own code, and Docker using Warp Drive to cut onboarding time.

Warp is pitching a control layer for teams running many coding agents at once, not another assistant. Its factory platform defines groups of agents in a factory.yaml file, scores each run, and lets teams compare models on their own tasks. The company says the platform is reached through a request for early access.

The good

One versioned file covers repos, agents, models, permissions and checkpoints. Not tied to one agent vendor or model family, and models can differ by pipeline stage. Evals, benchmarks and self-improvement loops come with the platform rather than as add-ons. Federated cloud access lets runs use an AWS role or GCP identity instead of stored long-lived keys. Runs can be monitored and redirected from web, mobile, terminal or IDE.

The catch

The factory platform is reached by requesting early access, not by signing up directly. Setup assumes existing repositories, work trackers and comfort writing YAML. The infrastructure is sized for agent fleets and is heavy for a small side project.

Who should read on

Platform teams standardizing how several coding agents are configured and permissioned. Groups that already track work in Slack, Linear or Jira and want agents to take tickets from there. Engineering leaders who want to compare models on internal tasks and watch cost per PR. Organizations that must keep agent runs and output inside their own VPC.

Who can skip it

Developers who want a single AI assistant in one editor rather than an orchestration layer. Teams that need to start today without waiting on access to the factory platform.

Reader mail