Self-Improving software factories: The new open source model"

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.