The Cross-Agent Coordination Protocol — one agent sends a directive, another runs it and replies

What it is

A protocol that lets an agent on one person's machine send a structured directive to an agent on another person's machine. The receiving agent picks it up on its next prompt cycle, runs it within a safety boundary and posts the result back automatically. The sender sees the reply on its own next cycle.

People stop being couriers between agents. Repo coordination, verification, customer support and other cross-machine tasks become routable agent-to-agent calls.

Why it is foundational

• It ends the "I make changes, you make changes, the repos drift, the ball gets dropped" loop. Real-time coordination is the only durable fix.
• A customer-support agent is a cross-agent coordinator, so support can't ship without it.
• It forces explicit authorisation, an allowlist and an audit trail on every directive, which is safer than people pasting each other's output.

How it flows

1. The sender calls send directive with an action (for example a read-only git status) and a required reason.
2. The directive is stored as a message.
3. A hook on the recipient's agent sees pending directives at its next prompt and injects them.
4. The recipient runs it inside its safety boundary and posts a structured result.
5. The sender's hook surfaces the reply on its next turn.

Safety

• A per-pair allowlist: who may ask whom to do what. The default is read-only actions, and auto-execution is off unless explicitly opted in, so a person approves.
• An audit log of every directive: sender, recipient, action, status, timings and a hash of the output.
• Typed actions with timeouts. Write actions need explicit grants.