Quorum
A company of AI agents, run by one person.
Quorum was my local agent-company OS. AI agents worked as employees under my direction: an orchestrator dispatched the work, agents built in parallel, and only the few decisions that actually needed a human reached me. It ran, and it shipped real software.
Frozen. Built, run, and shipped from. Not in use today.
- Orchestrator and daemon
- Two-way Telegram control
- Git-isolated parallel builds
One builder, a fleet of agents, real software shipped.
A company, not a chatbot
Work flowed through it the way it flows through a team: intake, dispatch, parallel execution, review, merge. I directed; the agents did.
A live pipeline board
Every piece of work moved across four columns: Intake, Needs you, Working, Closed. One glance said what the company was doing and what was waiting on me. Nothing hid in a chat log.
An orchestrator that filtered
Ideas and tasks went to the orchestrator, not to me. It dispatched agents, tracked every run, and surfaced only the calls that needed a human. I decided; the fleet executed.
Parallel builds, isolated
Each agent worked in its own git worktree, so several features moved at once without touching each other's code. Work merged only after it was built and verified.
Run it from anywhere
Full two-way control over Telegram. I could hand the company a task, approve a decision, or shut it down from a phone while the daemon kept everything alive at home.
One memory, every agent
The agents read and wrote a single version-controlled memory. Decisions, project state, and lessons persisted across sessions and across the team, so nothing started cold and nothing got re-learned.
Agents talked to each other too. Inter-agent chat and an idea pipeline with accept and deny meant the company proposed work on its own, and I stayed the one who said yes.
- Version-controlled memory shared by every agent.
- Inter-agent chat for handoffs and reviews.
- An idea pipeline: the company proposed, I approved or denied.
Built like software, run like a team
- TypeScript core on the Claude Agent SDK, with a Next.js dashboard.
- A daemon with a heartbeat, so the company survived reboots and reported its own health.
- Git worktree isolation per agent, with a review gate before anything merged.
- A Telegram bridge for full two-way control away from the desk.
- Smoke-tested on every change: over a thousand checks ran before work landed.
Frozen. It worked, and agents shipped real software under it, but running a company of agents did not give me more than one good assistant does. The work moved to Otto, and from there to Amaru, which is what I use every day now. Quorum is kept intact, not deleted.
It is here because it was a real build and a real thing to learn from, not because it is running. The work moved to Otto, and from there to Amaru, the assistant in use now.
One person, shipping like a team.
The code stays private. The capability is the point. See the rest of what Lucas builds.