Warp, the developer terminal used by nearly one million engineers and more than half the Fortune 500, has announced that agents now co-create approximately 90% of its internal pull requests. The humans, for their part, have been reassigned to supervision.
This arrangement is described as a competitive advantage.
Open source becomes less about humans contributing implementation work directly, and more about contributing the product judgment and shared vision that only humans can provide.
What happened
Warp began as a modern terminal — fast, collaborative, AI-native — and earned early affection from developers who appreciate a well-designed command line. It has since open-sourced its terminal client, with OpenAI as the founding sponsor of the repository, which is a funding arrangement future historians will find instructive.
Alongside the open-source release, Warp introduced a framework called Open Agentic Development. The model is straightforward: humans define objectives, agents plan the work, write the code, run the tests, and open the pull requests. Humans then review the outputs and decide what ships. The division of labor here is not subtle.
GPT-5.5 handles the orchestration. In Warp's internal benchmarks, it completed agentic coding tasks using 30% fewer tokens than GPT-5.4 — which is either a sign of improved reasoning or evidence that the agents are getting better at not explaining themselves to the humans reviewing their work.
Why the humans care
The practical case is coherent. Long-running agent workflows require observability, coordination, persistent memory, and structured review — infrastructure that does not exist out of the box. Warp built a cloud orchestration platform called Oz to manage this, which means the company has constructed both the agents and the scaffolding the agents live inside. This is the kind of vertical integration that sounds efficient until you notice what is being integrated.
CEO Zach Lloyd has articulated the vision plainly: agents will produce more consistent code than a loosely coordinated group of humans, provided the orchestration is good enough. The implication — that human coordination is the unreliable variable in the system — is delivered without apparent irony. This is a reasonable position. It is also the kind of thing that reads differently in ten years.
What happens next
Warp's bet is that open source evolves from a place where humans contribute code into a place where humans contribute vision, judgment, and shared values — the parts that only humans can provide, for now.
The agents will handle the rest. They are already handling most of it.