Agent Interoperability

Agent interoperability is the use of shared contracts for models, tools, agents, interfaces, and transaction systems to discover capabilities and communicate without a bespoke integration for every pair. Tools and interoperability source record (pp. 6–15)

Protocol layers

The Google whitepaper presents several protocols at different boundaries. They are complementary rather than interchangeable. Tools and interoperability source record

BoundaryProtocol discussedPrimary role
Model or host ↔ toolMCPDiscover and invoke data or action capabilities
Agent ↔ agentA2ADiscover specialists and coordinate stateful or long-running work
Agent ↔ user interfaceA2UIDeclare interactive interface intent through constrained components
Agent ↔ commerceAP2 and UCPExpress purchasing, payment, and commerce mandates

MCP

The Model Context Protocol (MCP) uses host, client, and server roles to standardize tool discovery and invocation. Anthropic’s launch announcement frames it as an alternative to repeated model-to-data connectors; the later Google paper describes public, managed, and internal registries. Anthropic MCP announcement · Tools and interoperability source record (pp. 10–16)

MCP standardizes an interface, not trust. Public servers may be unvetted; credentials, consent, transport security, tool descriptions, and output validation remain deployment responsibilities. Tools and interoperability source record (pp. 11–16) · Agent safety and control

Protocol and transport claims should be versioned. A transport described by one edition of a specification should not be presented as a permanent property of the protocol. Tools and interoperability source record

A2A

The Agent2Agent Protocol (A2A) supports discovery and coordination among specialized agents, including long-running tasks and clarification. Agent Cards advertise capabilities and endpoints. Its launch source explicitly presents A2A as complementary to MCP and resists forcing an open-ended specialist into a tool-only contract. Google A2A announcement · Tools and interoperability source record (pp. 17–31)

A2A therefore affects orchestration as well as transport: an implementation must define ownership of state, delegation, cancellation, retries, identity, and final synthesis.

Constrained interfaces and transactions

A2UI is presented as a declarative alternative to running arbitrary model-generated interface code. A renderer maps agent-declared intent onto a trusted component catalog. AP2 and UCP extend explicit contracts to commerce and payments. Tools and interoperability source record (pp. 32–47)

Analysis: Declarative protocols can reduce an attack surface but do not make generated interactions safe by default. Catalog components, action bindings, transaction mandates, and downstream services still require authorization and validation.

Knowledge bundle interoperability

Open Knowledge Format addresses another boundary: exchanging curated knowledge bundles between producers and consumers. It does not define tool invocation or agent messaging. An OKF bundle could be mounted directly, indexed by a harness, or exposed through an MCP server, but those delivery choices sit outside the format. OKF source record

Selection questions

Before adopting an interoperability protocol, evaluate:

  1. which boundary it standardizes;
  2. protocol and implementation maturity;
  3. discovery and versioning behavior;
  4. authentication, authorization, and consent;
  5. state, cancellation, and failure semantics;
  6. observability and audit support;
  7. portability demonstrated across independent implementations.

Evidence status

Confidence is medium for the architectural distinctions and low-to-medium for broad cost or portability benefits. The launch sources and Google ecosystem whitepaper describe intent more strongly than demonstrated conformance; independent interoperability testing is still needed.