← Field Notes

Field note · Protocols · July 2026

A2A v1.0: the enterprise boundary between agents

A2A does not replace MCP. It creates a separate contract for independent agents to discover one another, coordinate long-running work and exchange results across organisational boundaries.

AudienceCTOs · founders · engineering leaders
PerspectiveProduction architecture and delivery
FormatOriginal analysis grounded in primary sources

Why this matters now

Agent platforms are moving beyond one framework and one organisation. A2A v1.0 aligns agent interaction with web-friendly patterns such as HTTP, gRPC, polling, streaming and webhooks. That makes familiar gateways, identity, observability and scaling patterns relevant again.

The architecture boundary

Use MCP inside an agent to access tools and resources. Use A2A between agents when the remote capability remains an opaque, independently operated system. This distinction prevents every remote specialist from being reduced to a giant tool definition.

Production design

Treat Agent Cards as discoverable contracts, not trust. Add identity, tenant boundaries, capability allow-lists, task deadlines, idempotency, trace propagation, artifact validation and compensation for partial failure. Long-running A2A tasks should be durable workflows rather than fragile request chains.

My engineering perspective

The same principles used in event-driven order orchestration apply: stable contracts, correlation IDs, retries, dead-letter handling, replay and operational ownership. Agent interoperability is new; distributed-systems failure modes are not.

Executive takeaway

A2A creates optionality across vendors and platforms, but open communication without governance creates a larger attack and failure surface. Adopt the protocol together with an enterprise agent control plane.

Primary reading

  1. A2A Protocol : Announcing Version 1.0
  2. A2A and MCP : official comparison
  3. A2A Protocol Specification