Warp Oz: Multi-Harness Cloud Agent Orchestration and Cross-Session Agent Memory

Oz Becomes the First Multi-Harness Control Plane for Cloud Agents

Warp has significantly upgraded Oz, its cloud agent orchestration platform, with capabilities that fundamentally change how teams manage AI coding agents. As of May 20, 2026, Oz operates as the first truly multi-harness control plane: a single unified platform where teams can orchestrate Claude Code, Codex, and Warp Agent running concurrently, with consistent governance and audit logs across all of them.

The core premise behind multi-harness support is that no organization should be forced to bet its entire workflow on a single AI coding agent. Different agents excel at different tasks, and as the ecosystem evolves rapidly, teams need the flexibility to compare effectiveness across harnesses, route specific tasks to the most capable agent, and avoid being locked into any one provider. Oz now provides that infrastructure layer.

Automatic Multi-Agent Orchestration

One of the most significant additions is Automatic Multi-Agent Orchestration. For complex, long-horizon tasks — large feature builds, cross-codebase migrations, production deployment workflows — Oz can now spawn multiple subagents in parallel and coordinate them automatically. A management interface tracks progress across all active subagents in real time, with the ability to steer or intervene as work proceeds.

Previously, orchestrating multiple agents required manual coordination. With this update, Warp handles the task decomposition and parallel execution automatically, dramatically expanding the scope of work that agents can tackle in a single session.

Agent Memory: Cross-Session, Cross-Harness Knowledge

Warp introduced Agent Memory (currently in research preview), a persistent knowledge system that allows Claude Code, Codex, and Warp Agent to retain organizational context across sessions, repositories, and projects. Rather than starting from scratch every time, agents can now remember how a team structures its code reviews, what a system deployment topology looks like, and which conventions matter for a given codebase.

The memory corpus is owned by the organization — not Warp — and can be populated from files, databases, MCPs, or enterprise applications. Agents improve continuously as they index new knowledge from each session. This addresses one of the most persistent friction points with AI coding agents: the loss of context between sessions that forces engineers to re-explain the same organizational knowledge repeatedly.

Flexible Deployment and Session Continuity

Oz now supports self-hosting via Kubernetes, Docker, or direct execution within existing development environments, giving teams control over where their agents run. Session handoff capabilities allow work started on one device or environment to be continued seamlessly elsewhere — for example, initiating an agent task on a laptop, moving it to a cloud instance for a long-running job, then returning to a local terminal to review and iterate.

Who This Is For

Multi-harness orchestration is available to all Oz users while in beta. Agent Memory is in research preview with limited access. These capabilities are most immediately valuable to engineering teams managing multiple AI coding agents across a shared codebase, but the underlying architecture — persistent memory, parallel orchestration, and cross-harness flexibility — represents the direction Warp is building toward for all users of its agent platform.