AI Coding Agents Co-Create 90% of Warp PRs as Costs Surge

Agents Move From Assistant to Author

Warp, the company that built its name on a fast, AI-native terminal, says agents now co-create roughly 90% of the pull requests produced by its own engineering team 3. The figure appears in a case study published by OpenAI, which is the founding sponsor of Warp's newly open-sourced terminal client repository 3. Warp describes its approach as "Open Agentic Development." Humans set objectives and supervise results, while agents plan the work, write code, run tests and open pull requests 3.

The same period brought a very different disclosure from JetBrains. The IDE maker says its internal AI development spending rose about tenfold in six months, largely without anyone noticing at first 1. Read together, the two stories describe both sides of agentic coding: it is becoming the default way some teams ship software, and it is getting expensive enough that companies are building infrastructure just to see what they are paying for.

Warp's Pitch: The Terminal as Agent Workbench

Warp's reasoning is that the terminal already sits where commands, context, collaboration and review come together, which makes it a natural place for developers to work alongside agents 3. The company says nearly 1 million developers use its product and that it reaches more than 56% of the Fortune 500 3. A separate account puts the number at over 1 million active developers 4. The gap is small, but the two descriptions don't fully match, and "active" is a stronger claim than "used by."

That second account also says Warp open-sourced both its terminal and its Oz cloud-agent orchestration platform in April 2026, with financial backing from OpenAI 4. It places Warp as a fourth major player in the agentic terminal market, next to Claude Code, Codex CLI and Gemini CLI. On this reading, the "terminal won" argument now reaches past the large AI labs to independent tool makers 4.

The 90% number needs some context. "Co-create" is loose wording that could cover anything from an agent drafting a full change to one handling a single piece of it. It also comes from a vendor case study published by the company whose models Warp promotes. What may matter more is what Warp says it learned from running agents at that scale. Long-running agent workflows need observability, coordination, memory and human review before they can grow 3. Those are operational problems, not model problems, and they point to where the next round of tooling competition is likely to happen.

JetBrains: The Bill Comes Due

JetBrains' account shows the cost of that shift. The company says most of its developers use three to five AI tools a month, mixing CLI agents, agentic development environments for running several agents in parallel, and IDE-integrated assistants. They still use JetBrains' own IDEs as heavily as before 1. According to reporting on the post, token consumption began climbing sharply in January 2026 2. One observer called the 10x figure a rare quantification of the AI spending behind the company's losses, and noted JetBrains admitted it initially didn't know how to control those costs systematically 4.

JetBrains chose not to cap its engineers at one or two approved tools, even though it says many companies it talks to have done exactly that 1. Its argument is that the best option changes too quickly. One week it's Claude Code on Opus, then Codex, then something else 1. Instead, the company built a shared access and accounting layer meant to keep tool choice open while giving teams visibility into consumption and control over it 2.

What It Adds Up To

One story is upbeat and the other is cautionary, but they point the same way. Agents are moving from novelty to the main way some teams produce code, and the organizational overhead is growing with them. Warp's list of needs (observability, coordination, memory, review) and JetBrains' need for centralized cost accounting are two versions of the same problem. Companies running many agents across many tools need a control plane.

This suggests the competitive advantage in agentic development may move away from whoever has the best model this week. JetBrains' own experience implies that leadership won't last. The more durable advantage may belong to whoever governs the workflow, meaning who tracks spending, who keeps context across sessions, and who decides when a human signs off. Warp is betting the terminal is that place. JetBrains is betting on a neutral layer that sits above whatever tools its developers choose. Both bets assume agents are here to stay, and neither company is offering much evidence yet that the costs will level off.