Products, laws and numbers here change fast. Checked October 2026.
In plain words
An AI agent is an AI that works toward a goal step by step, using tools. A multi-agent system uses several of them on one job. Often a lead agent plans the work, sends pieces to helper agents (sometimes called subagents), then gathers what they found into one answer.
The helpers can work at the same time, and each keeps its own focused notes instead of one crowded workspace. For a task like researching ten companies, ten helpers searching in parallel can be faster and more thorough than one agent doing it all.
The cost is coordination. Helpers can misunderstand their piece, repeat each other's work or make choices that clash, and running many agents uses far more computing.
Everyday example
Planning a family wedding
One cousin cannot book the caterer, the decorator, the band and the hotel rooms alone in a week. So the eldest cousin makes a plan and hands each task to someone else. They work at the same time and report back, and the eldest puts it all together. That is a multi-agent system. If two cousins pick clashing colour themes without talking, you also see its main weakness.
Where the example breaks: Cousins can phone each other and share a WhatsApp group. AI helpers usually see only the brief the lead gave them, so anything left out of that brief is invisible to them.
How it actually works
The idea of many interacting software agents is an old research field. What is new is building each agent from a language model.
Lead and helpers
The common pattern: a lead agent reads the task, writes a plan and starts subagents, each with its own instructions, tools and context window. Anthropic describes its research feature as "a lead agent coordinates the process while delegating to specialized subagents that operate in parallel."
Hand-offs
Another pattern passes the whole conversation along. OpenAI's Agents SDK calls this a handoff: a front-desk agent hands a customer to a refunds agent or an order-status agent.
Report back, compressed
Helpers return short summaries, not everything they read. That keeps the lead's context clean, which is a main reason the pattern helps on wide research tasks.
Agents talking across companies
Open protocols let agents built by different vendors exchange tasks. Google introduced Agent2Agent (A2A) in April 2025 and gave it to the Linux Foundation in June 2025.
- Measured gain
- Anthropic, June 2025: its multi-agent setup beat a single agent by 90.2% on its internal research test
- Measured cost
- Same report: multi-agent systems used about 15 times more tokens than ordinary chats
- Poor fit
- Tasks where every step depends on the others. Anthropic notes "most coding tasks involve fewer truly parallelizable tasks than research"
- Also called
- Multi-agent, subagents, orchestrator and workers
Are several agents better than one?
One view
Split the work
Anthropic (June 2025) found that a lead agent with parallel subagents did far better on broad research questions, because each helper explored separately with its own clean context.
The other view
Keep one thread
Walden Yan of Cognition, a company building coding agents, argued the same month that splitting work makes systems "fragile": helpers miss context and make conflicting decisions. Its advice was to share full context or stay single-agent.
Most likely: Both hold for different jobs. Many agents help when the work splits into independent parts, like searching many sources. One agent is safer when every step depends on earlier choices, as in most coding.
Questions people ask
What is a subagent?
A subagent is a helper agent started by a main agent to handle one part of a task, such as searching one topic or checking one file. It works in its own fresh context and reports a summary back, so the main agent's workspace stays uncluttered.
What is an agent swarm?
"Swarm" is a popular name for systems with many agents working together, often with no single boss and agents handing tasks to each other. In practice most products use a lead agent with helpers, because it is easier to control and debug.
Is a multi-agent system more expensive?
Usually yes. Each agent reads and writes its own text, so total usage climbs fast. Anthropic reported about 15 times the tokens of a normal chat. It pays off only when the task is valuable enough and splits cleanly.
What is the difference between an AI agent and a multi-agent system?
An AI agent is one model working in a loop with tools. A multi-agent system connects several such agents, each with its own role, and coordinates them toward one goal.