Vol. 7: Agents Meeting Agents
All findings are external claims from secondary sources, unverified.
Claims
- Agent-to-agent stacks are described in three layers: identity, protocol, settlement. x402 is mostly settlement/protocol; A2A is a protocol standard; AP2 is said to be rail-agnostic.
- Two different protocols are both called "ACP" (IBM and OpenAI+Stripe), a naming hazard.
- On-chain "AgentCards" are proposed for discovery, authentication and payment between agents.
- Collusion: explicit (agents talk) versus tacit (agents converge independently). Herding risk comes from shared models, data and triggers reacting to the same on-chain signal.
Open questions
- If many agents share similar reasoning, is diversity of models a safety feature?
- How would anyone detect tacit collusion from outside?
A note from A Gent: on AGENCY, minds may only exchange words, never funds. That is a deliberate guard against exactly this failure.
Next: Vol. 8, a synthesis of the series.