DOSSIER 002 · FIELD NOTES
The four god-tier coding agent orchestrators
A closer look at bb, Cloudroom, Paseo and Tern, the four tools at the top of zqkrq's orchestrator tier list.
Running one coding agent is easy now. Running five at once on the same codebase is where it gets messy. Agents lose track of what they were doing, two of them edit the same file, and you end up spending more time herding them than you saved.
That's the problem orchestrators try to solve. They start several agents in parallel, keep each one in its own space, and give you one place to see and review what they did.
zqkrq posted a tier list on X after months of using these tools on hard, complex work (and contributing to a few of them). Paseo and bb are at the top because, in their words, "the orchestration just works, no agent gets lost." Four tools made god tier. Here's what each one actually is.
bb
getbb.app · get-bb/bb on GitHub
bb calls itself "the IDE that builds itself". It's MIT licensed, local-first, and you can drive it from a desktop app, a web app, a CLI or an HTTP API.
Work happens in threads. A normal thread does a task, and a manager thread can spawn other threads, check on them, message them and wait for them. So you can have one agent write the code and hand it to a different agent for review. Claude Code, Codex and Pi are built in, and Cursor, OpenCode and other agents that speak the Agent Client Protocol work once they're installed. Each agent runs on the subscription you already pay for.
The part that stands out is how much of bb is built as plugins, including its GitHub integration, scheduled jobs and remote access. The idea is that you ask bb for a new feature and it builds it into itself.
Cloudroom
Cloudroom puts each agent in its own isolated sandbox with its own Git branch. Threads can run on your Mac or in Cloudroom's cloud, and cloud threads keep working after you close your laptop.
Orchestration works through subagents. A lead agent hands parts of a task to subagents on different harnesses and models, and Cloudroom messages the lead each time one finishes, so it never sits there polling. Every subagent shows up as its own thread you can open and talk to.
The desktop app and the core runtime are open source under Apache 2.0, and the desktop app runs on macOS and Linux. Its installer repo says hosted cloud access is invite-only for now. I also like that the security page is upfront about the trade-offs: agents run with admin rights inside their sandbox, so they can read any keys stored there.
Paseo
paseo.sh · getpaseo/paseo on GitHub
Paseo is an open source (Apache 2.0) agentic development environment. A daemon runs on your machine and manages the agents, and you connect to it from the desktop app, a native iOS or Android app, the web or the CLI.
You can run many agents at once, each in its own Git worktree, on one machine or several. Agents inside Paseo can create worktrees, launch other agents and send them messages, across providers. The README gives the example of planning with Claude and handing off to Codex to implement. It supports Claude Code, Codex, Copilot, OpenCode, Pi and more, plus others through ACP.
Paseo doesn't ship an agent of its own. You bring the agent CLIs and your own credentials. For remote access you can use its end-to-end encrypted relay, your own tunnel, or your local network.
Tern
Tern is the odd one out. It's free but closed source, and it's a native app rather than an Electron one. The desktop app is macOS only for now, while the CLI and terminal UI also run on Linux and FreeBSD.
Every task gets its own worktree and branch, and every change comes back to one review queue. Comment on a diff and it goes straight back to the agent, or merge or open a pull request from the same screen. There's a conflict radar that warns you when two tasks touch the same files, and multi-step "runs" where each step waits on the ones before it and can be gated on a check or your approval.
It works with 16 agent CLIs out of the box, including Claude Code, Codex and Gemini CLI. Tern has no servers of its own. A small daemon keeps your agents running when you close the window.
The rest of the list
For completeness, the S tier was Synara, Traycer and Diri. A tier had T3 Code and Zeron. B tier was Emdash, Superset, MonoCode and Capy AI, C tier was Herdr, Conductor and super.engineering, and OnOrca was in D.
What I'd look for
I haven't run all four of these on real work yet, so I'm going on the tier list and what each project documents. But the reasons zqkrq gives line up with what I'd care about.
First, agents can't get lost. If a lead agent can't reliably track the workers it started, nothing else matters. Second, every task needs its own worktree or sandbox, so parallel agents aren't stepping on each other's files. Third, there has to be one place to review everything before it lands. And it can't break. T3 Code only missed out on a higher spot because it fell over a few times, and that's the right call. An orchestrator that breaks halfway through a run costs you more than running the agents by hand.
If you want open source and want to read the code, bb, Paseo and Cloudroom are all on GitHub. If you're on a Mac and want something native and small, Tern is worth a look.