Munder Difflin office of coding-agent clones running Codex and Gemini CLI in parallel

What they did

They installed Munder Difflin, the open-source Electron/React agent harness, and pointed it at the terminal coding CLIs they already pay for — OpenAI Codex and Gemini CLI — instead of buying new API capacity. Each CLI was registered as an agent on a shared office floor, where avatars at desks show what each clone is doing in real time. A coordinator clone (Michael) takes a task, splits it into assignments, routes messages between agents, and keeps a memory of what has already been done so work is not repeated. The agents run in parallel on the local machine, each in its own PTY-backed terminal session, so one clone can be drafting while another reviews and a third runs tests. Because the harness rides on existing subscription hourly limits, the office keeps working while the human is away, and the Command Center view is used to check progress, read live terminals, and step in when an agent stalls.

How it went

The setup produced a working parallel office: multiple clones active at once, visible on the floor, with the coordinator handing out and tracking assignments. Polish is pre-release (0.5.6), so expect rough edges, but the core loop of dispatch, message, remember and review held up across a multi-step task. The main win was throughput — several agents progressing simultaneously on one repo without the human babysitting each terminal.

Worth knowing

Pre-release software: expect breaking changes and occasional stuck sessions. It depends on the CLIs you already have installed and authenticated, and it consumes your subscription hourly limits, so heavy parallel use can hit rate caps. Local LLM and bring-your-own-key paths need extra configuration. Agent-to-agent messaging can loop or duplicate work if the coordinator prompt is vague, and there is no strong sandboxing story beyond what the underlying CLI provides — review diffs before merging.

Wiring

Install Munder Difflin (Electron app) and make sure codex and gemini CLI are installed and authenticated on the same machine. In the harness, add each CLI as an agent so it gets its own node-pty terminal session rendered with xterm.js. Designate one agent as the coordinator (the clone) and give it the routing and memory responsibilities. Agents communicate through the harness message bus rather than directly; the coordinator assigns tasks, collects results, and writes state to memory so restarts do not lose context. The Pixi.js office floor is just the visualization layer over those sessions — the Command Center is where you inspect and intervene. Keep concurrency modest at first to stay inside subscription hourly limits.

Try it yourself

Set up a Munder Difflin office in this repo using Codex and Gemini CLI as the two worker agents. Register both CLIs, confirm each can run a trivial command in its own terminal session, then create a coordinator clone whose job is to split a task into at most three parallel assignments, route messages between the workers, and keep a short memory file of completed work. Test it on this task: audit the repo for missing tests, have one agent write the tests, another review them, and the coordinator summarize the diff. Report what broke.