The Office, but your AI agents are the paper salesmen

What it does

Munder Difflin is an Electron desktop app that wraps real terminal-agent CLIs — claude, agy, and codex — into a multi-agent hive. Each agent runs in an authentic pseudo-terminal (node-pty), but instead of working in isolation, they share a file-based mail system, long-term memory, and a blackboard, all coordinated by a supervisor GOD agent that routes tasks and only asks for human approval on critical or destructive operations. You watch the whole thing unfold as pixel-art avatars wandering a 2D office floor.

The interesting bit

The project leans fully into the Dunder Mifflin parody: the orchestrator sits in “Michael’s office,” agents walk to stations, and messages fly desk-to-desk as animated envelopes. Under the pastel SNES aesthetic is a surprisingly serious engineering effort — per-agent git worktrees, a cost circuit breaker with a “steer → constrain → stop” ladder, OpenTelemetry observability, and a markdown-first semantic memory layer that claims to be the “fastest memory layer in the world.”

Key highlights

• Real multi-provider hive: Claude Code, Antigravity (Gemini), and OpenAI Codex agents collaborate through a shared on-disk layer of inboxes, outboxes, and append-only logs.
• Human-in-the-loop governance: the GOD agent resolves routine traffic autonomously and queues only spend, destructive ops, or scope changes for your approval.
• Heavy operational tooling: per-agent token budgets, live cost telemetry, context-window gauges, a circuit breaker for runaway agents, and durable SQLite persistence across restarts.
• Task management with dependencies: a kanban board assigns work to agents and respects task order, plus recurring scheduled missions with an adaptive heartbeat.
• External integrations: GitHub issue ingestion, CI status watching, Slack/webhook ingress via tunnelmole, and desktop notifications when agents finish or stall.

Caveats

• The repository wears a “working prototype” badge, so polish is explicitly not guaranteed.
• The README calls its memory layer the “fastest memory layer in the world” without offering benchmarks or numbers.

Verdict

Worth a look if you’re already running multiple terminal AI agents and want a local, visual coordination layer with real operational guardrails. Skip it if you need a polished production tool or a single-agent experience — this is explicitly a prototype, and its value is in the ensemble.