One brain for the whole fleet.

Claude Code, Codex, Cursor, Gemini and Grok on one memory, with gates between them and a different agent checking every maker.

For: Someone running several coding agents who is tired of re-explaining everything to each.

First win

Two agents, one task, one reviewer, one receipt. The maker never reviews its own work.

Proof: A multi-agent run with a handover note and an eval receipt.
Compounds as: A fleet ledger of routing decisions, costs and failures.

Every first win ends in a receipt. See the receipt format.

Where this role stands

Concept

No form, no checkout and no delivery date. The first win is not yet tested with a real person.

What already exists

SIS, swarm, command, ops hub

Starlight-Intelligence-System
starlight-swarm
starlight-command
agentic-ops-hub
workflow-tier-plugin
agents-council
prompt-engine

Fleets spend your own tokens.

Two names on a receipt

Illustration

Three agents work from one memory. One writes the change. A different one reviews it and rejects it, with a reason. The first fixes it and the second approves. The receipt carries both names and both verdicts. Nothing reached main because one agent liked its own work.