PWSM: an orchestration command center for teams of AI agents

PWSM is the system that makes everything else we do possible. It is how DDD runs teams of AI coding agents at the same time, instead of one assistant at a time, and it is why a small senior studio can ship like a much larger one.

The challenge

A single AI assistant is a bottleneck. It works one task at a time, and the human becomes the message bus between it and the rest of the work. Running several agents at once sounds obvious, but it is dangerous: agents that edit the same files at the same time create merge chaos. We needed parallelism that was actually safe.

What we built

PWSM, the Parallel Worktree Session Manager, is a command center for orchestrating AI agents. It spawns and coordinates many agents concurrently, each one working in its own isolated git worktree inside a sandbox, so they never collide. It exposes a control layer of orchestration tools and a library of specialist roles, plus a live web command center: a task board, a dependency graph, a real time event feed, and terminals for every agent.

File level conflict detection and an auto finalize watchdog keep the parallel work clean and reviewable, so output from many agents lands on the main branch without stepping on itself.

How we built it

We built PWSM the way we build for clients: agentically, in tight loops, with a human holding the plan and the review. It runs on Node, Express, and WebSockets, with node-pty and xterm.js for the live terminals, Docker for isolation, and the Model Context Protocol to expose its tools to AI clients.

The result

PWSM is the machine behind our delivery promise. When we say we build in an agentic manner, this is what we mean concretely. It is the reason the highest impact build lands in your first weeks rather than your first quarter, and it is a working answer to the question every buyer should ask: show me that you actually build this way, not just talk about it.