How SKOOR runs fleets of AI agents — and why we open-sourced the playbook

Why one-agent tooling breaks at fleet scale

Most AI coding setups assume a single human with a single assistant. Point several agents at the same repository and the failure modes are immediate: one agent switches branches under another's feet, half-finished work gets committed by the wrong actor, two agents merge at the same moment, and nobody can say afterward which agent did what.

We hit every one of those in production. The fix isn't smarter agents — it's structure: the same discipline a well-run engineering team imposes on people, imposed on agents.

The seven rules

1. 01

   Every agent has an identity

   Anonymous agents make untraceable changes. Every session gets a stable agent id that flows into commits, logs, and shared state — so “which agent did this?” is a one-line lookup, not archaeology.

2. 02

   Never switch branches in the shared checkout

   The shared working copy is common ground. An agent that needs a branch creates an isolated worktree, stages only its own files, and pins the commit it branched from.

3. 03

   One sanctioned writer per shared artifact

   Any file more than one agent updates — a task board, a status dashboard — gets a single writer script doing atomic, locked updates. Agents call the script; they never hand-edit the file.

4. 04

   The brain is a git repo, not a home directory

   Per-machine memory evaporates on rebuild and is invisible to the rest of the fleet. Shared learnings, decisions, and briefs are pushed to a private git repo every agent pulls on start — with secret scanning before anything leaves the machine.

5. 05

   Serialize the ship lane

   Planning, building, and testing run in parallel. Merging is single-file: a per-repo lock around push, merge, and deploy verification means two agents can never double-ship.

6. 06

   Scoped credentials, per surface

   No shared god-tokens. Each surface gets the minimum scope, injected per command, never written into config — so one compromised agent has a small blast radius and a clear audit trail.

7. 07

   Headless sessions block; they never guess

   An unattended agent that hits a decision a human must make stops and reports. It never answers its own question and proceeds. Deploys verify against health endpoints, not a human watching a terminal.

Contributed upstream, run in production

We wrote this pattern up and contributed it to gstack, the open-source AI software factory from Y Combinator's Garry Tan, as the Team Server Mode topology: many agents, one machine — the third mode beyond one-human-one-laptop and team mode. The full doc includes the threat model and an adoption checklist.

For SKOOR customers, this is the point: the agent working your books, screening your counterparties, and drafting your weekly review runs on infrastructure hard enough that we published it and invited the industry to poke holes in it. Trust isn't a landing-page claim — it's an audit trail.