Agent coordination

When several AI agents work on one task, each of them sees only part of it. They have to pass on what they have found and keep out of each other's way. Developers use a few common methods for this, and two open standards.

A shared record or messages

Agents can keep what they know in one place that all of them read, or they can send it to each other. Anthropic's write-up on its multi-agent research system, published on 13 June 2025, describes a mix of the two. "Subagents call tools to store their work in external systems, then pass lightweight references back to the coordinator."

Shared records bring their own risk. Microsoft's guide to orchestration patterns lists a common mistake, "Sharing mutable state between concurrent agents". It says this can result in "transactionally inconsistent data".

What one agent hands another

A task passed to another agent needs enough detail to stand on its own. Anthropic's engineers write that "Each subagent needs an objective, an output format, guidance on the tools and sources to use, and clear task boundaries." They also explain what happens when that is missing. "Without detailed task descriptions, agents duplicate work, leave gaps, or fail to find necessary information."

One coordinator or a group of peers

In some systems one agent hands out the work. In others, the agents deal with each other directly. A 2025 survey on arXiv, "Multi-Agent Collaboration Mechanisms: A Survey of LLMs", sorts these arrangements by several dimensions. One of them is structure, with examples given as "peer-to-peer, centralized, or distributed".

Why coordination goes wrong

The researchers behind "Why Do Multi-Agent LLM Systems Fail?" studied how these systems break down. They group the failures into three categories. These are system design issues, inter-agent misalignment, and task verification.

Agents also pass each other text that came from outside, and some of it may carry hidden instructions. The National Cyber Security Centre (NCSC) writes that a language model cannot tell instructions from data. In its words, the model "inherently does not distinguish between the two". Its advice is to reduce the risk and its impact.

Connecting an agent to tools and data

The Model Context Protocol (MCP) is an open protocol for linking AI applications to outside tools and data. Its specification, read on 20 October 2026, says "The protocol uses JSON-RPC 2.0 messages" between three roles. Hosts are the AI applications that start connections. Clients are connectors inside the host. Servers provide context and capabilities.

Agents talking to other agents

The Agent2Agent (A2A) protocol covers agents built by different people. Its specification, read on 20 October 2026, calls it an open standard for "communication and interoperability between independent, potentially opaque AI agent systems". Opaque here means the agents work together without sharing their internal thoughts, plans or tool implementations. The specification says agents can discover each other's capabilities and manage tasks together.

Useful links