The biggest real-world obstacle to multi-agent collaboration is often not technology but trust: letting one agent 'call' another means deciding what gets exposed and what stays secret. A2A answers with 'Preserve Opacity' — collaboration does not require transparency. Agents exchange a 'minimal contract' via the Agent Card: what I can do, how to connect, and which interaction modes I support — nothing more. The internal state machine, memory structure, proprietary algorithms, and tool implementations all stay inside the black box. This design solves two problems at once: 'can cooperate' — as long as both speak A2A they interoperate; and 'dare to cooperate' — an enterprise doesn't have to hand over core assets for cross-org collaboration. Decoupling 'capability' from 'implementation' is a key step toward cross-company agents.
For Zero to Pro Agent Harness learners, 'decouple capability from implementation' is one of the most worth-chewing principles in agent architecture: a well-designed agent should be able to state 'what I can do' without opening 'how I do it'. This maps to the course's core on external interface design — the interface is the contract, the implementation is private. For Zero to Codex learners, you can draw the analogy to Codex plugins and MCP integrations: when you wire in a tool, you care about its exposed schema and capability, not its internal implementation. Understanding 'opaque collaboration' helps you judge whether a multi-agent scheme is 'elegantly delegating' or 'dangerously sharing internal state'.