Many business leaders assume that if one AI agent can handle a task, several agents working together must do it even better. New research from Vals AI challenges that view. In controlled evaluations, teams of AI agents barely outperformed solo agents while costs rose by up to 5.1 times. For organizations already watching AI spending, that is a clear signal to pause before building elaborate multi-agent workflows. The findings are especially relevant for remote and hybrid teams, where AI experimentation can spread quickly without central oversight.

When More Agents Do Not Mean Better Results

The study compared individual frontier models against multi-agent teams using models including GPT-6 Sol and Claude Opus 5.5. Across four test scenarios, only one produced a measurable quality gain from teamwork. The other three delivered essentially the same output as a single agent while consuming far more tokens and compute. In practical terms, a task that costs one amount with a solo agent could cost more than five times as much when assigned to a team, with no meaningful improvement in accuracy, completeness, or usefulness.

This runs counter to the popular assumption that adding agents mimics human collaboration. In theory, one model can draft, another can review, and a third can fact-check. In practice, agents often duplicate work, pass redundant context back and forth, and generate internal text that does not change the final answer. The expense comes from coordination, not from better reasoning. A single well-prompted agent is frequently all that a defined task requires.

Not every task is the same. The isolated case where a team improved suggests that some complex workflows may benefit from parallel perspectives, particularly when tasks can be cleanly divided and independently checked. But even then, the gain has to be weighed against the higher bill. The burden of proof should sit with the team advocating for extra agents, not with the manager trying to contain costs.

The Hidden Cost of Agent Coordination

Multi-agent systems are not inherently flawed, but their overhead is real. Every additional agent reads prompts, produces responses, and shares information with other agents. Those exchanges consume tokens at each step, which translates directly into higher API bills and longer response times. When the task is relatively contained—such as summarizing a document, classifying a request, or drafting a routine message—a solo agent with clear instructions is usually sufficient. Adding reviewers, critics, and planners can simply multiply token consumption.

Finance and operations leaders should therefore treat multi-agent architecture as a hypothesis rather than a default. A practical approach is to run a small controlled comparison. Give one agent a defined task and measure quality, latency, and cost. Then test a multi-agent version on the same task. If the team setup does not produce a clear improvement in output, the simpler system is the rational choice. This kind of pilot testing can prevent AI budgets from ballooning while delivering no better service to employees or customers.

What This Means for Workplace AI Adoption

For HR teams, IT departments, and people managers, the research strengthens the case for disciplined AI adoption. The goal is not to maximize the number of agents an organization can deploy, but to solve real problems at a reasonable cost. In recruiting, a single agent can screen resumes, extract skills, or draft interview summaries. A multi-agent pipeline in which one model reviews another model's notes may slow the process and raise costs without improving candidate outcomes.

HR leaders can also apply this lesson by auditing current AI tools. Many teams already pay for assistants, copilots, and agent features that overlap. Consolidating to one reliable tool for a specific job may reduce both subscription costs and employee confusion while making outputs easier to audit.

The same logic applies to distributed work. Remote teams often adopt AI tools at different speeds, and complex agent workflows can spread inside a company before their value is proven. The Vals AI findings encourage teams to separate enthusiasm for agentic AI from the harder work of measuring impact. Before scaling a tool, leaders should ask whether it reduces time, improves quality, or lowers cost—and whether a simpler setup could do the same.

For organizations using platforms like XMF to coordinate flexible talent and remote operations, the lesson is relevant: clearer task design and stronger human coordination often beat adding more automated layers. AI agents can be valuable, but value comes from fit, not from the number of agents involved.

Originally published by XMF, inspired by publicly reported industry news.