Zach Lloyd

Alt titles: Agent orchestration with message passing / Agent orchestration for every model / Warp’s approach to agent orchestration With models getting more capable, we’ve quickly scaled from single agent problems to multi-agent problems – How can agents delegate tasks to accomplish ever-larger goals? You may have heard of “agent swarms” or “agent teams” in this arena, but they come with drawbacks: model lock-in, complex UX, or both. We want to share how we’ve tackled orchestration with our model-agnostic platform, Oz. Our approach has some unique goals: - Support any model, *and* any harness (`claude`, `codex`, etc) - Delegate across local instances *and* across isolated cloud sandboxes - Provide a UX that requires zero tmux or TUI knowledge to use We’ll explore how we implemented message passing across harnesses, how we handle agent sandboxing with Docker containerization + serverless deploys, and how we designed these primitives to make a system that works with any agent. You’ll walk away with a clear outline of how to build agent orchestration well. Plus, we invite you to try our Oz orchestration platform and tell us what you think. Talk format: Primarily a tech demo and code walkthrough. We’ll show multiple examples of tasks that are best served by delegation, and show both local and cloud-based runs. We’ll also walk through the design of our message passing implementation at a high level to show how it works.

Most teams start with one AI coding agent on a single developer’s machine. But scaling that work across large, complex codebases and process automation tasks is difficult in large organizations. How do teams move agent workloads to the cloud? What happens when different tools like Claude Code and Codex are used across teams? How do agents coordinate and share what they’ve learned? At Warp, merged PRs are created by agents running on Oz, Warp’s cloud agent infrastructure. In this session, we'll show how to move from a single local agent to the cloud, and fan out tasks across multiple harnesses (Claude Code, Codex, Gemini CLI) coordinated from a single orchestrator. The session demonstrates multi-harness management, multi-agent orchestration, and Agent Memory keeping context intact across sessions. We'll talk about the challenges of building Oz and how this agent infrastructure helps accelerate Warp’s customers and its engineering organization. 3:11 PM