How Warp Uses Claude to Power an Agentic Developer Terminal

The terminal is one of the oldest interfaces in software development, and for most of its history it changed very little. When Warp launched in 2021, the founders saw an opportunity to modernize the experience—better UI, smarter completions, and eventually AI capabilities as language models became capable enough. The company added natural-language-to-shell-command translation, a chat assistant panel, and eventually “agent mode,” among the first command-line agents with direct terminal access.

As agentic coding accelerated in 2024, teams trying to deploy autonomous coding agents hit a new problem: the terminal wasn’t built for it. Running agents in the cloud meant stitching together environments, observability, steering tools, and handoff mechanisms for local work. Warp’s existing product handled local developer tasks well, but the infrastructure for production-grade agentic workflows didn’t exist.

The turning point came when Claude Sonnet 4 arrived. Up to that point, Warp’s terminal agent had excelled at “shallow but broad” tasks—translating natural language commands, navigating CLIs, and helping developers find the right syntax. With Sonnet 4, the agent could handle the full software development lifecycle: writing code, running tests, reviewing diffs, and debugging across an entire codebase. Warp’s engineering team noticed the step-change immediately during internal testing. For coding tasks, the platform’s auto mode now defaults to Claude, with Opus, Sonnet, and Haiku selected based on task complexity.

Warp built Oz, a cloud agent orchestration platform that lets users launch Claude Code and Warp Agent runs directly from the terminal or via integrations with Slack, Linear, and GitHub. Teammates can observe active agent sessions and intervene with the right permissions. The platform supports self-hosting, Warp’s managed infrastructure, or a hybrid of both. Warp Agent has logged 40 million conversations total, with 65% of those tokens routed to Claude. The terminal has run 10 million Claude Code sessions in total, with over 400,000 each week.

The reach of the platform has also surprised its builders. Non-technical users—marketers, analysts, operations staff—are seeking out Warp as an entry point into internal CLIs and data tools they couldn’t otherwise access, with the agent handling the unfamiliar parts. Warp’s next phase centers on persistent cross-agent memory in the Oz platform, so agents working across different harnesses can carry context forward from one task to the next.