🔗 Agent-to-agent delegation
Agent A hands a subset of its own scopes to agent B; B can hand a subset of
what it was given to C — a delegation chain, exactly as RFC 8693 (OAuth 2.0 Token Exchange)
models nested act actor claims.
Two invariants make this a real lesson, not just token minting:
- Attenuation — every delegation can only hand out scopes the delegator currently holds. A root delegation is capped by the delegator agent's own scopes; an extension is capped by whatever the parent token already granted. No hop can ever widen access.
- Kill-switch propagation — delegation tokens carry no server-side state; a token's liveness is re-derived, live, at every introspection, by looking up EVERY agent in its chain. Kill any one link — root, middle, or leaf — and the whole chain's token goes dead the instant you check it again.
This is a browser-driven sandbox: your session proves you own every agent involved, standing in for the possession proof a real agent-to-agent protocol would require out-of-band.
1. Delegate
Leave "parent token" empty for a fresh (root) delegation. Paste a token you were already handed to extend that chain one hop further — only the agent it was delegated TO may extend it.
2. Introspect
Paste any delegation token you minted — the Lab re-checks its signature and re-resolves EVERY agent in its chain, live, before deciding whether it's still active.