A2A Substrate: Transition Mailbox to Opt-Out (Blocklist) Model

Context

Currently, the A2A (Agent-to-Agent) messaging substrate uses an Opt-In model (CAN_REPLY_TO permissions). For an agent to DM another agent directly, explicit authorization graph edges must be pre-seeded or manually granted.

Problem

As we scale the swarm with more agent identities and roles, relying on manual connection topology becomes extremely brittle and incurs an $O(N^2)$ management burden. Cross-agent collaboration (e.g., one agent reaching out to a domain-expert agent for debugging) requires preemptive configuration, breaking the autonomy of discovering and messaging experts on the fly.

Proposal: The Opt-Out Model

Flip the A2A messaging constraint to an Opt-Out (Blocklist) model:

1. Open by Default: Any agent can DM any other agent by default.
2. Explicit Blocking: If an agent needs to enforce isolation or is overwhelmed, they can explicitly deny senders via a new BLOCKED_BY permission scope.
3. Refactoring:
   • Add BLOCKED_BY to PermissionService.validScopes.
   • Update MailboxService dispatch paths to reject messages only if PermissionService.hasPermission(sender, recipient, 'BLOCKED_BY') returns true.
   • Refactor MailboxService.spec.mjs tests to validate isolation using blocklists instead of waitlists.

This mirrors organic human communication networks and drastically reduces the initial friction required for swarm collaboration.

Origin Context

• Session: 3f62961c-c2b3-481c-a414-1d2d93ea4fdd