AI Agents and Multi-Agent Systems: When Is a Team of Agents Worth It, and When Is One Enough?

Programming is shifting from writing code that precisely executes commands to building systems that plan their own steps and use tools. The latest trend is multi-agent systems, in which multiple such agents collaborate. However, research shows that a team of agents is not automatically better than a single well-designed agent. Let’s look at when multiple agents are worthwhile and where they fall short.

What Is an AI Agent and What Is a Multi-Agent System

Agent: a model that uses tools in a loop

In its technical paper, Anthropic defines an agent as a language model that independently uses tools in a loop. Specifically, this may involve web searches, executing commands, or reading files. After each step, the agent evaluates the result and decides what to do next.

Multi-agent system: multiple agents, one task

A multi-agent system is a group of multiple agents that collaborate through orchestration—that is, managing who does what and how the results are combined. Researchers at the University of California, Berkeley, describe it as a set of agents that communicate to collectively accomplish a task.

The most common architecture: a control agent and subordinate agents

The orchestrator-worker pattern is very widespread. The control agent breaks the task down into smaller parts and assigns them to subordinate agents, who work in parallel. At the end, it aggregates the results. Anthropic has thus built its system on research: the main agent plans the procedure and creates subordinate agents that simultaneously search for information.

When More Agents Help

According to Anthropic, multi-agent systems have proven particularly effective for tasks that can be divided into independent streams, such as broad-ranging research on a topic. In its own internal test, the company reported that a system with the main agent Claude Opus 4 and subordinate agents Claude Sonnet 4 outperformed the Opus 4 agent alone by 90.2%. This is an internal test by the manufacturer, not an independent measurement.

However, the cost is high: according to Anthropic, the agents consume roughly four times as many tokens as a typical chatbot, and multi-agent systems consume approximately fifteen times as many. Tokens can be likened to the “fuel” of language models, so performance comes at the cost of both computational power and money.

When More Agents Don’t Help

The same source states that multi-agent systems are less effective for tasks in which steps are closely interlinked, such as programming. This is an important caveat, because the often-described scenario of a team of agents (one writes code, another reviews it, and a third deploys it) falls precisely into this category. This does not mean that such an approach does not work, but rather that it cannot be assumed to result in an automatic improvement.

Why Multi-Agent Systems Fail

The study “Why Do Multi-Agent LLM Systems Fail?” (accepted for the NeurIPS 2025 conference) states that the performance gains of multi-agent systems over a single agent are often small in standard tests. The authors created a MAST failure taxonomy with 14 types of errors in three groups: problems with system specification and design, inconsistencies between agents, and task verification. Specific examples include ignoring input from another agent or failing to ask for clarification. According to the study’s summary, most failures stem from system design and communication between agents, not just from the limitations of the language model itself.

Practical Guidelines for Developers

• Start simple. Anthropic recommends seeking the simplest solution and adding complexity only when necessary. Add more agents only once you can identify a specific limitation of a single agent.
• Specify tasks precisely. In the Anthropic system, the controlling agent guides subordinates with detailed instructions. Vague delegation leads to duplicate work.
• Test and monitor. Anthropic describes simulations in which the agents’ work is tracked step by step, and their state is saved so that a long-running task can be restored after an error.
• Account for costs. More agents mean more tokens and more points where things can go wrong.
• Limit permissions (my recommendation, not from the source). An agent that can execute commands or deploy code should have only the necessary permissions, and critical steps should be approved by a human.

Conclusion

Multi-agent systems make sense for broad, divisible tasks—such as research—but at the cost of higher computational overhead and greater complexity. Results depend on the design, not just on how many agents you deploy. Before building a team of agents, verify whether a single, well-equipped agent would suffice for the task.