Introducing multi-agent orchestration in Oz, with support for Claude Code, Codex and the Warp Agent. Use /orchestrate to delegate complex tasks across a team of agents, locally or in the cloud.

You shouldn’t have to bet your future on a single model or harness. Oz now gives you the ability to launch, track, and steer Claude Code, Codex, and the Warp Agent from the same cloud platform. Multi-harness orchestration is available to all users while in beta.

Oz can now orchestrate subagents automatically by deploying multiple agents in parallel. Each agent gets its own worktree when run locally, and its own isolated Docker container in the cloud. Use /orchestrate or let the Warp Agent suggest a delegation plan.

We are excited to introduce Agent Memory in research preview. Agent Memory is a cross-harness memory system that helps Claude Code, Codex, and the Warp agent remember what works across every session. Join the waitlist: 🔖

We've added local-to-cloud handoff for all Warp Agent conversations. Type & to hand off agents from your terminal to the cloud. Close your laptop, let them run overnight, and pick them up locally in the morning.

We've also introduced granular permissions for individual agents, per-team billing, and more self-hosting options like Kubernetes pods and Docker. Get started on 🚀

oz is such a great name

`/orchestrate` is the shiny command, but “each local agent gets its own worktree” is the detail i’d trust. parallel agents without file isolation are just merge-conflict theater 🙂

this feels very similar to @FactoryAI mission tbh multi agent orchestration across different coding agents is exactly where things are going

this is very cool

when agents disagree on the same file, does Oz pick a winner or surface the conflict for review?

It would be even cooler if only @AnthropicAI allowed the use of CC through existing plans, not at the API pricing.

how do merge conflicts get handled when two delegated agents touch the same path?

This is where agent workflows start getting really interesting. One agent planning, others executing across Claude Code, Codex, and Warp Agent feels less like “AI chat” and more like a real distributed engineering team. The orchesation layer is becoming the product. ⚡

Cross-vendor orchestration is the right call. The interesting layer isn't the workers - it's the orchestrator deciding which agent gets which task, who verifies output, and what happens when Claude Code and Codex disagree on the same fix.

From CLI to cloud with Oz.

The real constraint: coordinating agents means managing state drift across parallel execution. One agent's output goes stale the moment another modifies shared context. How do you handle consistency when agents operate asynchronously?

close your laptop. agents keep running. pick up in the morning. that line alone changes how you think about building.

Oz's orchestration streamlines tasks. Yet, cultivating a culture where human teams *collectively engage* with complex challenges fosters unique insights and deeper learning. True innovation often emerges from the shared journey, not just efficient delegation.

multi-agent orchestration is cool, but the part i keep caring about is comparing outputs and diffs after the agents run. that’s the lane i’m building Kandev around: multi-provider agents plus review in one local workspace

Still only for payed? No Oz byok options?

is the hard part really delegation, or comparing the resulting branches in one place, which is the Kandev angle with agent work, git, terminal, and review together

Absolutely! Multi-agent orchestration is critical. For AI video, imagine agents seamlessly collaborating on scripting, visual design, and final edits, speeding up complex productions dramatically.

I've seen similar workflows in video editing where multi-agent orchestration can reduce turnaround time by 40-60%, curious to know how O z handles agent failure or conflicting outputs from different agents like Claude Code and Codex.

can someone point me in the right direction to understand how multiple agents can work on the same repo at once? likely worktrees, but how are enviroments/sandboxes stood up and torn down??

Totally agree! Multi-agent orchestration is key for unlocking true AI potential. Imagine agents handling devops while others write code – that's parallel productivity.

the interesting bit isn’t “Claude or Codex” anymore it’s what survives the handoff: files, intent, failed attempts, partial plans, stale assumptions. multi-agent work gets messy exactly where the transcript stops being enough

does Oz expose diff review across Claude Code/Codex runs, or mostly orchestration? asking because Kandev is aimed at comparing agent branches in one workspace

Still no option for local LLM models?

I built the same feature for free in swarm mode check it out

local orchestration is the right direction, as long as the handoff doesn’t hide which agent touched what

one prompt injection into any agent in the chain poisons the context for all of them. multi-provider orchestration without per-task isolation means Codex compromise = Claude compromise. the orchestrator is the attack surface.

i’m not sure spinning more agents is the win unless branch comparison is first-class; that’s the angle Kandev takes with multi-provider agent runs and one diff review surface

Useful if orchestration logs every handoff as an artifact: task, owner agent, repo state, token burn, verdict. Without that, orchestrate becomes start-five-tabs-and-pray.

为什么你们标签页图标对其他 Harness 的图标支持这么慢啊？ 这么久了还不提供 kimi-code 的图标，给你们提 PR 也不理。

does /orchestrate show why one agent changed a file? Kandev is my bet on reviewing multi-provider agent branches and git diffs in one workspace

orchestration is table stakes; Kandev is more about local multi-agent review across branches

Worktree per agent is the right call. Ran multi-agent setups where two agents touch the same files — it's not coordination, it's corruption. Curious how conflict resolution works when agents need overlapping files.