Multi-Agent Orchestration Patterns And Agent-To-Agent Protocols

Once one agent works, the next idea is usually to add more: a researcher, a writer, a checker, each with its own instructions and tools. Splitting work this way can help, but it also adds cost, delay and new places for things to go wrong. The design of the split matters more than the number of agents.

This article describes the patterns that model developers document most often, shows how agents coordinate through a manager or through handoffs, and explains where the Agent2Agent (A2A) protocol fits alongside the Model Context Protocol (MCP) as of October 2026.

Start With One Agent

OpenAI’s guide recommends getting as much as possible out of a single agent before adding more, because extra agents bring extra complexity and overhead.¹ It suggests splitting only when there is a clear sign of strain: instructions full of if-then branches that are becoming hard to manage, or a set of tools so similar that the agent keeps picking the wrong one even after the descriptions have been improved.

Anthropic, writing in June 2025 about its own multi-agent research system, added a limit from the other direction. At that time it judged tasks where every agent needs the same context, or where the pieces depend heavily on one another, to be a poor fit for several agents. Most coding work, it noted, has fewer truly independent parts than open-ended research does.²

Five Building-Block Patterns

Anthropic’s engineering guidance names five patterns for combining model calls.³ It presents all five as workflows, with the overall shape set in code, although in some of them the model decides much of the detail, such as how to split a task. The same patterns also turn up as parts of larger agent systems.

Two Ways To Coordinate Agents

OpenAI describes two broad shapes for systems with several agents.¹ In the manager pattern, one central agent treats the others as tools. It calls a specialist, receives the answer and stays in charge of the conversation and the final result. In the decentralised pattern, agents act as peers and pass the work along. OpenAI defines a handoff as a one-way transfer: when one agent hands off, the receiving agent takes over execution and is given the latest state of the conversation.

The choice comes down to who needs to see the whole picture. A manager suits work that must be combined into one answer, such as a report drawn from several sources. Handoffs suit a sequence of specialists, such as a customer service flow where a triage agent passes a billing question to a billing agent, which then deals with the customer directly.

What Extra Agents Cost

Anthropic’s research system is a well-documented example of the orchestrator pattern. A lead agent plans the investigation and starts several subagents that search in parallel, then combines what they find into an answer. In June 2025 Anthropic reported that this setup, with Claude Opus 4 leading and Claude Sonnet 4 subagents, outperformed a single Claude Opus 4 agent by 90.2 percent on its internal research evaluation. It also reported that agents typically use about four times as many tokens as a chat exchange, and multi-agent systems about fifteen times as many. Tokens are the small chunks of text a model reads and writes, and they are the usual basis for what a model costs to run.² These are the vendor’s own measurements on its own test, not independent results.

The useful lesson is the trade-off rather than the headline. Several agents can explore more ground at once, but each one consumes model capacity and adds coordination. A multi-agent design makes sense when the gain clearly outweighs that spend. Work that splits into independent parts benefits most, because those parts can run in parallel; sequential handoffs can still be worthwhile, but they buy clarity of roles rather than speed.

How Agents From Different Vendors Talk

MCP standardises how one agent reaches tools and data. A separate question is how an agent built by one company finds and delegates to an agent built by another. That is the job of A2A.

Google announced A2A on 9 April 2025 as an open protocol designed to complement MCP. Each agent publishes an Agent Card, a document in JSON (a common structured data format) describing what it can do, so that a client agent can pick a suitable partner and send it work.⁴ The project moved to the Linux Foundation, and IBM’s Agent Communication Protocol merged into it in August 2025. Version 1.0, the first stable specification, was released in March 2026 and added signed Agent Cards so that an agent’s identity can be checked cryptographically. On 17 August 2026 A2A became a hosted project of the Agentic AI Foundation, which also hosts MCP.⁵ In April 2026 the Linux Foundation reported that more than 150 organisations supported the protocol, a figure supplied by the foundation.⁶

Opening an agent to others raises questions of trust, identity and permission that go beyond orchestration. Those are covered in Agentic AI Security. The next article looks at why multi-agent systems fail and which controls make them dependable.