Orchestration Standards

AI agents are only as useful as the tools and data they can access. Without standards, every integration is custom: each framework invents its own way to connect to databases, APIs, and external services. This leads to duplicated effort, fragile integrations, and agents that can't work together across vendor boundaries.

Two complementary standards have emerged to solve this. MCP (Model Context Protocol) standardises how agents connect to tools and data sources. A2A (Agent-to-Agent Protocol) standardises how agents communicate with each other. Together, they form the interoperability layer that makes multi-agent systems practical at scale.

MCP was announced by Anthropic in November 2024 as an open standard for connecting LLMs to external tools and context. Think of it as USB-C for AI agents: a single, standardised interface that any agent can use to connect to any tool.

An MCP server exposes capabilities — tools (functions the agent can call), resources (data the agent can read), and prompts (templates the agent can use). An MCP client (the agent) discovers these capabilities and invokes them through a standard JSON-RPC protocol. The server can run as a local process, a remote HTTP service, or even in-process within the agent runtime.

As of April 2026, MCP has native support in LangGraph, CrewAI (v1.10.1+), OpenAI Agents SDK, Vercel AI SDK, Claude Agent SDK, and Mastra. This means any tool exposed as an MCP server is immediately usable by agents built on any of these frameworks — write once, use everywhere.

The MCP roadmap for 2026 focuses on four areas. First, transport scalability: keeping the set of official transports small but evolving existing ones to handle production workloads reliably.

Second, agent communication: handling longer-running work with async Tasks, so an agent can kick off a job, do other work, and retrieve results later. This is critical for support workflows where a tool call might take minutes (like running a diagnostic script on a customer's account).

Third, enterprise readiness: audit trails, SSO-integrated authentication, gateway behaviour for routing and rate limiting, and configuration portability so MCP server configs can be shared across teams.

Fourth, ecosystem growth: making it easier for tool vendors to publish MCP servers and for teams to discover and compose them. The goal is a rich ecosystem of pre-built integrations, similar to how npm packages or Docker Hub images work today.

A2A was originated by Google in April 2025 and has since been donated to the Linux Foundation for open governance. While MCP connects agents to tools, A2A connects agents to each other — enabling multi-agent collaboration across vendor and framework boundaries.

The protocol uses a client-server model. A client agent formulates a task and sends it to a remote agent. The remote agent executes the task and returns results. Agents advertise their capabilities through Agent Cards — JSON documents describing what the agent can do, what inputs it expects, and what outputs it produces.

A2A supports multi-modality (text, audio, video streaming) and is designed for both synchronous and asynchronous collaboration. An agent built on CrewAI could delegate a subtask to an agent built on LangGraph, without either system knowing or caring about the other's implementation.

MCP and A2A are complementary, not competing. MCP is vertical: it connects an agent downward to the tools and data it needs. A2A is horizontal: it connects agents sideways to other agents that have different specialisations.

In a customer support scenario, imagine a triage agent that receives a ticket. It uses MCP to read the customer record from the CRM, query the knowledge base, and check the order status. Based on its analysis, it uses A2A to delegate the response drafting to a specialised reply agent running on a different framework. The reply agent uses its own MCP connections to access tone guidelines and response templates, then returns the draft via A2A.

This separation of concerns — MCP for tool access, A2A for agent collaboration — is what makes the standards practical. Each solves one problem well, rather than trying to be a universal protocol for everything.

For support teams evaluating AI tooling, the adoption of MCP and A2A has three practical implications.

First, reduced vendor lock-in. If your agent framework supports MCP, you can swap frameworks without rebuilding all your tool integrations. The MCP servers you've set up for CRM access, knowledge base search, and order management continue to work regardless of which agent framework calls them.

Second, composability. You can combine best-of-breed components: use one vendor's triage agent, another's reply agent, and your own custom escalation logic, all communicating via A2A. This is especially valuable for large organisations that may have different teams managing different parts of the support workflow.

Third, future-proofing. As both standards reach production maturity in 2026, building on them means your agent infrastructure is aligned with where the ecosystem is heading. Tools and agents that speak MCP and A2A will have access to the widest range of integrations and collaboration partners.