Group identity, charters and attestation for AI agent swarms

Running many agents as one accountable unit

Every identity standard names the agent. None names the group. When an orchestrator delegates to three workers and one spawns two more, every question an operator or auditor asks is about the six together. This paper gives the group a name the server mints, a charter that is enforced, a budget that conserves, a stop whose propagation is measured and whose limits are stated, and a 24-hour signed attestation a counterparty can verify offline.

• Eight design rules for governing a group of agents, each backed by shipped code
• A table of eleven controls with what each one bounds and, more useful, what it does not
• What "stopped" can honestly mean, with the three verdicts that keep a timeout from becoming a success
• Two numbers to ask any vendor for: kill-switch propagation and attention utilisation
• The one gap we have not closed, stated plainly

1. Executive summary
2. Why a swarm needs a name of its own
3. What we built, and what each control does not bound
4. Stopping a swarm: what "stopped" can honestly mean
5. Eight design rules for governing a group of agents
6. Two numbers to ask any vendor for
7. What we do not claim

6 pages, 18 min read. Published for Cybersecurity Awareness Month, October 2026. May be shared and quoted with attribution.