Multi-agent setups are currently considered an obvious next step in automation: instead of a single language model handling a task, several specialised agents share the work, hand over intermediate results and check each other. The expectation behind this is simple: more perspectives, fewer errors, better results. An analysis by testing provider Vals AI calls this expectation into question.

The result of the investigation: teams of AI agents delivered barely better results than individual agents, but caused up to 5.1 times higher costs. Only in one of four test runs was any measurable advantage shown at all. The testing was carried out with current models, including GPT-6 Sol and Claude Opus 5.5.

Where the costs arise

The cost driver in such setups is tokens – the text units into which the inputs and outputs of a language model are broken down and according to which they are billed. Each additional agent brings its own context: system instructions, task description, previous intermediate states, results from the other agents. This context is processed again at every step.

This means that the effort does not grow linearly with the number of agents. Coordination, follow-up questions and repetitions generate additional text traffic that has to be paid for – regardless of whether it improves the quality of the solution. According to the analysis, data from Anthropic also points in the same direction: from around ten agents, quality stagnates while token costs continue to rise.

Why more agents are not automatically smarter

The idea that agents correct each other presupposes that the errors arise independently of one another. However, when several instances of the same or a similar model work on a task, they often share the same blind spots. A second agent is then more likely to confirm a false assumption than to expose it.

Added to this is the loss of information at the handover points. What one agent passes on to the next is a condensed summary – details, uncertainties and reasoning are partly lost in the process. The more handovers a task goes through, the greater the risk that small inaccuracies add up to an incorrect overall result.

Consequences for development projects

For companies planning AI-supported development or process automation, this does not mean doing without agents, but it does mean asking questions in a different order. Instead of asking how many agents a workflow needs, it is worth asking which subtask actually justifies an independent instance.

In practice this means:

  • Start with one agent. A single, well-instructed agent with clean context is the baseline for comparison. Only when it demonstrably fails to solve a task should splitting it up be considered.
  • Parallelisation only with genuine independence. Multiple agents are worthwhile above all where subtasks can really be worked on separately – such as simultaneously searching different data sources – and not where they constantly have to wait for each other.
  • Measure cost per result. The relevant metric is not the token total per month, but the effort per resolved ticket, generated module or reviewed document.
  • Check quality against a fixed reference. Without defined test cases, it is impossible to assess whether a more complex setup actually works better or merely differently.
  • Solve deterministic steps conventionally. Validations, format checks or database queries belong in regular code, not in another agent.

Architecture beats mass

The more interesting lever is usually not the number of agents, but the quality of what an agent has available: precise tools, reliable interfaces to its own systems, compact and up-to-date context. An agent that can access a well-maintained product database or a cleanly documented API directly generally delivers better results than three agents passing incomplete summaries to each other.

This also includes a sober look at model selection. Not every work step needs the strongest available model. A division by difficulty – simple classifications on smaller models, complex analyses on more powerful ones – often lowers costs more significantly than any change in the number of agents.

Assessment for web and software projects

For web and software projects, this means: a multi-agent setup is an architectural decision with ongoing operating costs, not a quality feature in itself. Anyone integrating AI into development, editorial processes or data maintenance should take the simplest working structure as a starting point and justify every additional instance with measurable benefit. What ultimately determines the result is clean data, clear interfaces and robust test cases – precisely the foundations that decide the success of a project even without AI.

Sources