2026-04-22 · 4 MIN READ
Swarms that ship
Three rules for running a small army of AI agents end-to-end without burning the house down.

There is a version of agent orchestration where you sit and watch a prompt loop forever and call it work. Skip that version. The shape that ships looks like a small org chart: one coordinator, a handful of specialists, tight role boundaries, and a checkpoint after every task. Agents that report nothing get nothing back. Agents that commit progress to shared memory get rehired.
Rule one: phase the work. Specification, pseudocode, architecture, refinement, completion. Each phase has a tight contract and a single output the next phase can consume. Skipping phases is faster on the way down and slower at the bottom of the well.
Rule two: humans review the diffs, not the chatter. The interesting artifact is always the patch. If you find yourself reading agent transcripts to understand what changed, the agents are writing the wrong things to memory.
Rule three: keep the gaming card out of the AI lane. Pin the AI work to the cards you can spare and leave the one driving your monitor alone — a job that never fights a game for the GPU finishes without either one stuttering. Boring infrastructure, boring deploys, interesting work in the middle. That is the whole arc.
- 01claude-flow
Coordinator and hive-mind tooling powering the swarm.
https://github.com/ruvnet/claude-flow
- 02
- 03SPARC methodology
Specification → Pseudocode → Architecture → Refinement → Completion.
https://github.com/ruvnet/claude-flow#-sparc-development-methodology
- 04A swarm that builds games
Where these rules went: intent to a playable app, no keyboard touched.
/blog/swarm-that-builds-games
- 05The hive reported false-done
The failure these rules guard against, and the gate that caught it.
/blog/hive-false-done