Agent2Agent (A2A)
Agent2Agent is the agent-to-agent interoperability layer: MCP exposes tools to an agent; A2A lets independent agents discover capabilities, negotiate modalities, and coordinate tasks without sharing private state. Source: A2A protocol specification, https://a2a-protocol.org/latest/specification/, 2026-06-17
A2A is useful when the unit of delegation is an agent, not a tool call. The protocol frames remote agents as opaque systems with discoverable capabilities, supported input/output modalities, task state, and security boundaries. That makes it a coordination primitive for cross-vendor or cross-org agents where direct access to memory, tools, or chain-of-thought should stay private.
The wiki should treat A2A as complementary to MCP as the Integration Standard, not as a replacement. Use MCP when Kevin needs an agent to call a tool, database, browser, service API, or local capability. Use A2A when Kevin needs one agent system to delegate work to another agent system and preserve each side's autonomy.
In Kevin's ideation stack, A2A pairs with Services-as-Software: A2A is how specialized agents communicate; services-as-software is the business lens for what those agents sell, namely outcomes rather than dashboards.
Routing Rules
| Need | Prefer |
|---|---|
| Call a tool or service from one agent | MCP as the Integration Standard |
| Coordinate two independent agent systems | A2A |
| Build a typed agent harness inside one app | Tool-Calling Design and app-native APIs |
| Delegate a business outcome across vendors | A2A plus Services-as-Software |
Protocol boundary
A2A is for agent-to-agent delegation, not ordinary function calls. That distinction matters:
| Layer | Unit of work | What remains private |
|---|---|---|
| Function/tool call | one action with schema | tool internals |
| MCP server | resources, prompts, and tools exposed to one agent | server implementation |
| A2A agent | task handled by another autonomous agent | remote agent memory, tools, chain-of-thought, internal policy |
The protocol's value is that each side can remain opaque while still advertising capabilities, input/output modalities, task state, and security requirements. In Agent Capability Registry terms, a remote agent becomes a governed capability.
Product implications
A2A needs control-plane support, not just SDK support:
- identity for the calling and called agent,
- capability discovery and trust metadata,
- task status and cancellation,
- auth and billing boundaries,
- artifact exchange,
- audit logs for both sides,
- fallback when the remote agent refuses or fails.
This is the remote-agent version of Agent Product Surface: the user should know which agent owns the task, what authority it has, and what evidence came back.
Concept Position
| Field | Value |
|---|---|
| Concept family | Agent harness and runtime primitives |
| Concept owned | Agent2Agent is the agent-to-agent interoperability layer: MCP exposes tools to an agent; A2A lets independent agents discover capabilities,... |
| Category map | Concept System Map |
Timeline
- 2026-07-01 | Concepts category refresh added this page to the Agent harness and runtime primitives family, linked it to Concept System Map, and kept it standalone because it owns this reusable mental model: Agent2Agent is the agent-to-agent interoperability layer: MCP exposes tools to an agent; A2A lets independent agents discover capabilities,... Source: User request, 2026-07-01
- 2026-06-18 | Cross-linked Agent Capability Registry because A2A agent cards are a remote-agent capability discovery surface. Source: whole-wiki graph audit, 2026-06-18
- 2026-06-19 | Expanded protocol boundary and control-plane implications: A2A makes remote agents discoverable capabilities, not local tools. Source: whole-wiki concept review, 2026-06-19
- 2025-04-09 | Google announced Agent2Agent as a protocol for secure agent communication and coordination across enterprise platforms. Source: Google Developers Blog, https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/
- 2026-06-17 | Added as a canonical concept page to resolve repeated A2A mentions across ideation, Loop, and services-as-software pages. Source: whole-wiki section audit, 2026-06-17