Warp cuts agent costs to $15 per pull request with factory model

Warp has pivoted to providing open infrastructure for what it calls cloud software factories, defined as agentic systems that build software with measurable autonomy. The platform requires these factories to be defined declaratively in code using YAML, specifying repositories, MCPs, skills, and workflows. This approach enables version control and self-improvement, as coding agents can submit pull requests to the factory definition itself. Warp centralizes agent conversations and artifacts from entry points such as Slack, Linear, Jira, and GitHub to create a closed-loop system for observability. The company supports benchmarking by replaying historical runs against different model and harness configurations, including Claude Code, Codex, and Pi, to optimize cost and correctness. While Warp uses Namespace as its primary sandbox provider, it supports pluggable compute layers like Daytona, Modal, or E2B, and allows custom inference providers to utilize existing commitments with Anthropic, OpenAI, or Google. The target market is companies with at least five to ten engineers, as smaller teams are advised to use direct subscriptions instead. Internally, Warp reports that its initial adoption of this approach cost approximately $80 per pull request due to inefficient model usage, such as using high-powered models for triage. Through tuning and model routing, the company reduced this cost to between $15 and $20 per pull request. Concurrently, Warp measured a 40% increase in pull request throughput over the last month without an increase in revert rates. In benchmarking examples, GPT-5-6-SOL achieved 95% correctness at an average cost of $6, while GLM 5.3 Flash achieved 93% correctness at an average cost of $2. Zach Lloyd, of Warp, asserts that direct API usage is 5 to 50 times more expensive than subscription plans, citing that a $200 plan provides $8,000 worth of API value. Lloyd argues that companies must move from intuitive approaches to measurement-based factory engineering to efficiently deploy coding agents, emphasizing that defining factories in code is foundational for self-optimization and cost control.