Warp is making a sharper infrastructure argument than most agent workbenches. The product is not just a nicer place to chat with a coding agent. It is becoming a control plane for fleets of long-running agents across multiple harnesses and multiple model backends.
What Changed
On May 19, 2026, Warp said Oz now acts as a multi-harness control plane for Claude Code, Codex, and Warp Agent. Warp says Oz can launch, track, govern, and steer those agents in one place, while also orchestrating subagents, managing cross-harness memory, and supporting self-hosted deployment options.
On May 20, 2026, Warp added bring-your-own inference, including custom endpoints compatible with the OpenAI Chat Completions API. That means a team can keep Warp's interface and agent layer while choosing its own provider, router, gateway, or internal inference stack.
The Operating Proof
OpenAI's May 27, 2026 customer story on Warp adds unusually concrete validation. According to OpenAI, Warp has nearly 1 million developers, is used by more than 56% of the Fortune 500, and sees around 90% of its own internal pull requests created with agents.
That is useful because it shows the product is not only describing an agentic future. It is already testing a version of it inside a real software organization with visible scale and throughput.
Why The Abstraction Matters
There are really two separations happening here. First, Warp is separating the control plane from any one agent harness. Second, it is separating the harness from any one model provider. Those are distinct layers, and companies will want optionality at both of them.
If that separation holds, a software organization can switch between Codex, Claude Code, or future harnesses without rebuilding supervision, permissions, memory, and workflow management every time. It can also change inference routing without giving up the agent interface and review surface users already know.
The ZHC Angle
This fits cleanly with themes we have covered in AI gateways, Stainless, and deployment companies. As the stack matures, more value moves into the layers that coordinate and govern agent work instead of the layers that only expose a model.
For zero-human company builders, that means the winning workbench may look less like an assistant and more like an operating console for persistent digital labor.
The Take
Warp's latest launches suggest that multi-agent software production is becoming an infrastructure design problem. The hard part is no longer just getting one agent to solve one task. It is coordinating many agents over time with enough optionality, memory, review, and governance to make the system usable inside a company.
Oz is one of the clearest product bets yet on that exact problem.
Related: See our prior notes on AI gateways, Stainless, workspace agents, and deployment companies.