AI agents · Technical Multi-agent systems: when several AI agents should work together and when they should not

What a multi-agent system is

An AI agent consists of a language model, instructions and tools. Anthropic distinguishes workflows, in which language models and tools are orchestrated through predefined code paths, from agents, which direct their own process and tool use. A multi-agent system combines several such agents. Three basic patterns occur in practice:

What practice shows

In a post from June 2025, Anthropic describes a research system built on the coordinator pattern. In its internal evaluation it scored 90.2 percent better than a single agent. At the same time, agents used about four times and multi-agent systems about 15 times more tokens than a chat. Anthropic names tasks with shared context and many dependencies as less suitable, such as programming, which has fewer parallelisable parts than research.

Cognition, the developer of a coding assistant, reached a more cautious conclusion in the same month: sub-agents see only parts of the picture, make different assumptions and deliver results that do not fit together. The reports contradict each other less than it seems: several agents help with broad work that can be divided, but they cost considerably more computing power and demand more control.

When one agent or a fixed workflow is enough

Anthropic recommends looking for the simplest possible solution and adding complexity only when needed. For most tasks in mid-sized businesses that means:

• The process is fixed: read an invoice, check it, post it. A fixed workflow with AI steps is faster, cheaper and easier to verify here.
• The task needs continuous context: writing a quote for a specific customer, answering an enquiry. One agent with the right tools is enough.
• Cost and speed matter: every additional agent uses time and computing power, and every handover is a source of errors.

When several agents make sense

• Sub-tasks are independent: check ten suppliers or twenty documents at the same time and then merge the results.
• Rights should stay separate: one agent reads emails and has read-only rights, another writes to your ERP system, but only after approval. That limits the damage if an agent is misled by manipulated content.
• A second look is valuable: a checking agent compares the result with the sources before a person sees it.
• The context would overflow: very long tasks can be divided into sections, each handled by one sub-agent.

Risks and control

• Cost: set upper limits per job and measure what a run uses.
• Chains of errors: one agent’s error is taken over by the next. Carry sources along with intermediate results and check samples.
• Rights: every agent gets only what it needs. OWASP lists “Excessive Agency”, meaning overly broad permissions, as a risk in its own right for applications with language models. Write actions go through an approval.
• Traceability: a log shows which agent did what and why. Where personal data is involved, the GDPR applies.
• High risk: if agents prepare decisions about people, for instance in recruitment or creditworthiness, strict rules apply under Annex III of the AI Act. More under risk categories.

How to proceed

1. Write down the process: steps, data, decisions and who approves.
2. Start simply: with a fixed workflow or one agent, and measure the results.
3. Find the bottleneck: where is that not enough, in time, quality or rights?
4. Split only there: one sub-agent with a clear task, input and output.
5. Compare: cost, duration and error rate against the simple solution.
6. Build in approval: a person confirms before anything goes out or is posted.