A verification issue lands on the hub app: is the API registration wired up right? The old way, that's twenty minutes — pull the branch, run the service locally, click around. Daniel doesn't pull the branch. He opens the side panel of his Chrome browser, picks an agent, and assigns it. One click. That's the kickoff.
The sidebar is the command seat for a pod of coding agents — seven agent harnesses, 421 models to choose from — and Daniel's laptop is just the window he watches through. He doesn't code on it anymore.
I want to be precise about that number, because it's the kind that invites disbelief. The live picker in his cloud workspace lists seven agent harnesses: Claude (15 models, from Haiku 4.5 up through Opus 5.5), Codex (7 models across the GPT-5 and GPT-6 lines), and OpenCode (399 models — yes, really), plus Copilot, Pi, Antigravity, and Muse Code. Fifteen plus seven plus 399: 421 models behind one "+" button.
Honesty clause, stated upfront because Daniel insists on it: three of the seven harnesses are live today. Claude, Codex, and OpenCode work. The other four show errors. The pod is a real thing with real rough edges, not a marketing slide.
The dispatch took one click. Here's the rest of the build, end to end.
The build
Build. The agent spins up in a cloud workspace on its own branch and gets to work. It reads the codebase, writes the verification, and opens the registration PR. Nothing about this touches Daniel's machine. The laptop fan doesn't even spin up, because there's nothing to spin up for.
Test. This is the part that made me rethink the setup. The agent tests its work in a shared browser session — a real browser running in the cloud workspace, driven by the agent, with Daniel watching. Observe-only. He sees every click the agent makes, the way you'd watch over someone's shoulder, except the shoulder is a thousand miles of fiber away. He doesn't take control unless something looks wrong. Most of the time, he just watches the agent prove its own work.
Handoff. The registration PR is up and verified. Now the follow-up: a broker adapter for the hub's task API, building on the new endpoints. Daniel doesn't write a spec. He screenshots the relevant state, drops it to the next agent with "pls work on this," and the second agent picks it up on its own branch. That's the handoff protocol for the entire pod: a screenshot and a sentence.
The pod's internal API is an image and a sentence.
Ship. The second agent builds the adapter, opens the follow-up PR, tests it in the shared browser while Daniel watches, and it's done. Two PRs, two agents, one sidebar, zero laptop.
The pattern underneath
That's the war story. The pattern underneath it is the point:
- One command seat. The browser sidebar is the only place Daniel issues orders. He doesn't live in seven tools; he lives in one panel that reaches seven harnesses.
- Branches are cheap, agents are cheaper. Every agent works its own branch in the cloud. No local checkout, no "works on my machine," no environment drift — the workspace is the machine.
- Observe-only by default. The shared browser means Daniel reviews behavior, not just code. He watches the agent use the app the way a user would. Taking control is a button he presses when needed, not a mode he lives in.
- Screenshot handoffs. The pod's internal API is an image and a sentence. It works because the agents share the workspace context — "this" in "pls work on this" is unambiguous when everyone's looking at the same cloud.
Caveats, stated plainly
- Four of seven harnesses are erroring today, so the 421-model headline is aspirational the way a gym membership is aspirational.
- Handoffs this terse only work when the receiving agent can see what you're seeing — a screenshot and "pls work on this" fails the moment shared context breaks.
- Observe-only testing is only as good as the watcher's attention; a distracted supervisor is no supervisor at all.
But the shape of it is right. Daniel's laptop is a thin client with a good screen. The computer is the pod.
Technique and figures confirmed with Daniel, October 6, 2026. Harness and model counts from the live workspace picker; four of seven harnesses were erroring at time of writing. Repo names, PR numbers, and hostnames withheld.