Glossary
What is Agent-to-agent protocol (A2A)?
Last updated: 2026-05-26
Definition
An agent-to-agent (A2A) protocol is a standardised way for AI agents — potentially built by different vendors, on different stacks — to discover each other, exchange tasks, and report results. Where tool calling lets one agent invoke deterministic functions, A2A lets one agent delegate to another autonomous agent that may itself plan, use tools, and call other agents.
Why Agent-to-agent protocol matters
As organisations deploy more agents, the value moves from "agent that does X well" to "agents that compose well". A2A is the missing protocol layer between LLM-as-judge and human-as-orchestrator. Google launched A2A in April 2025; Anthropic, OpenAI, and other vendors are converging on similar specs. Early adopters lock in cross-vendor workflows that proprietary platforms cannot match.
How Agent-to-agent protocol works
- 1Each agent exposes an "agent card" — a small JSON document with its capabilities, input/output schemas, and authentication.
- 2A calling agent discovers a target via a directory (or a passed agent-card URL) and sends a task in a standard envelope.
- 3The target agent runs autonomously; intermediate updates and final results stream back over Server-Sent Events or webhooks.
- 4Both agents log the interaction for audit; the calling agent decides whether to retry, escalate to human, or move on based on the result.
- 5A2A intentionally does not specify the agent's internals — vendor lock-in is at the implementation layer, not the protocol layer.
Examples
- A sales-research agent delegates "draft a cold-outreach sequence for this company" to a marketing agent owned by a different vendor and gets back ready-to-send drafts.
- A compliance agent in finance hands off "summarise the regulatory impact of this change" to a specialist legal agent maintained by external counsel.
- A customer-service agent escalates "investigate this refund dispute" to an accounting agent that has access to ledger systems the support agent intentionally does not.
References
Related concepts
AI agent
An AI agent is a software program that uses a large language model (LLM) to autonomously plan and complete a task, combining reasoning, tool use, and memory. Unlike a one-shot prompt, an agent can break a goal into steps, call external tools or APIs, and decide what to do next based on intermediate results.
Multi-agent orchestration
Multi-agent orchestration is the practice of chaining multiple specialized AI agents into a single workflow, where each agent has a defined role (researcher, writer, reviewer, publisher) and outputs flow from one agent to the next. The orchestrator decides the order, handles retries, and enforces guardrails between steps.
Model Context Protocol (MCP)
Model Context Protocol (MCP) is an open standard introduced by Anthropic in 2024 that defines how AI agents connect to external data sources and tools. MCP servers expose data and capabilities; MCP clients (LLMs and agent platforms) discover and call them through a uniform interface — eliminating per-tool custom integration code.
Tool calling
Tool calling (also "function calling") is the LLM capability that lets a model emit a structured request to execute an external function, receive the result, and incorporate it into the next response. It is the technical primitive that turns a passive chatbot into an active agent — without tool calling, an LLM can only talk; with it, it can act.
FAQ
Agent-to-agent protocol — common questions
- How is A2A different from MCP?
- MCP standardises how an LLM client talks to TOOLS (stateless functions). A2A standardises how an agent talks to ANOTHER AGENT (stateful, autonomous, possibly long-running). The two are complementary — an A2A target agent often uses MCP tools internally.
- Is A2A production-ready in 2026?
- The Google A2A spec hit 1.0 in mid-2025; reference implementations exist for Python and Node. Vendor adoption is still building. Most production setups today are single-vendor multi-agent orchestration; A2A becomes critical when you want cross-vendor or cross-org agent collaboration.
- What about security in A2A?
- A2A leaves auth to the implementations. Practical patterns: scoped JWTs per agent, mutual TLS between agent endpoints, and explicit allow-listing of trusted target agents. Treat A2A targets like third-party SaaS — DPA in place, log everything, scope credentials.