The Strongest Teams of AI Agents Will Be Built Using Different Models

The agentic workforce is on the march.

AI agents—highly autonomous AI systems that can understand context, make decisions, and carry out complex actions—now routinely operate alongside human workers in many industries. Software developers increasingly call on coding agents for tasks such as writing, testing, and reviewing code. In customer service, AI agents support human agents in call centers, triaging customer queries, summoning information, and suggesting resolutions. In supply chain management, human network planners can turn to teams of AI agents that monitor supply and demand conditions, propose plans, optimize inventory, and orchestrate activity between customers and suppliers.

Agentic AI is so prevalent that some business leaders now count AI agents as part of their regular workforce. In a recent episode of HBR IdeaCast, McKinsey & Company Global Managing Partner Bob Sternfels observed that his company workforce now numbers 60,000, of which 20,000 are AI agents. That’s up from just 3,000 agents 18 months prior. NVIDIA CEO Jensen Huang has an even more expansive vision—that NVIDIA someday will be a 50,000-employee company with 100 million AI assistants in every single group.

Yet, without attention to the diversity of this burgeoning agentic workforce, business leaders are likely to find that many of the promised benefits of agentic AI—increased productivity, innovation, creativity—fail to fully materialize. In fact, an emerging body of research now points to significant performance improvements from diversity in agentic systems. One study, for instance, showed that agent teams selected with diversity in mind were 25% better at resolving software engineering problems than agents acting individually, mainly due to the blending of different skills and knowledge sets. Another study showed that just two diverse agents can “match or exceed the performance of 16 homogeneous agents.” The message is clear: like diversity in human workforces, agentic diversity pays significant performance dividends.

While the evidence around the case for agentic diversity continues to mount, it is less clear how business and technology leaders can actually create diverse agentic teams in practice. Is it enough to adapt existing agents to different cultural or social settings, or imbue them with distinct personalities? In short, what makes for diversity in agentic teams and how do you go about creating it?

Most enterprises currently appear unprepared for these questions. They are still largely focused on the technical and business aspects of AI implementation.  This article sets out the causes and consequences of lack of diversity in agentic AI systems, and the actions that leaders and technologists can take now to create more diverse agentic AI teams.

Agentic AI’s Diversity Challenge

In principle, large-language (or “foundation”) models, which represent the “brain” of an agentic system, can be prompted to create surface diversity in AI agents, in the form of different personality types, ways of thinking, and social and cultural attitudes. AI agents can, for example, be set “hot” or “cold” to mirror (human) extrovert or introvert personalities, and they can be directed to be more questioning, challenging, or conciliatory. But in many cases these changes are largely cosmetic.

I interviewed Enver Cetin, director at the AI company Ciklum, who articulated the underlying problem: “When clients talk to me about diversity in agentic AI, they usually mean personality or cultural diversity at the agent layer. The real problem I see across financial services, automotive, and retail is that nearly everyone is running on the same handful of foundation models, the same retrieval architectures, often the same data sources. When the stack underneath is uniform, dressing the agents in different personas is mostly cosmetic. Costume change is not cognition.”

This view is reinforced by a growing body of research highlighting the lack of diversity in AI models. One study has shown that prompting for different personality types leads to AI models that are very binary in their thinking and actions—very outgoing or introverted, for example—whereas most humans exhibit these traits on a continuum.

Another study by Atari et al has shown that the responses of major large-language models such as ChatGPT to psychological profiling tests resemble those “of people from Western, Educated, Industrialized, Rich and Democratic societies”—what they term “WEIRD” populations. By contrast, the major models fail to capture the diversity of other populations with very different values.

The Business Consequences of Non-Diversity in Agentic Teams

 The lack of underlying diversity in agentic AI models has far-reaching consequences for individual teams, businesses, and markets.

At the organizational level, agentic diversity matters because numerous studies have shown that personality and cultural variation are important determinants of team success, partly because they create “cognitive friction” that enables teams to solve complex problems faster. As AI agents becoming more integrated with human teams and achieve a greater preponderance in the workforce, the risks of stifling different perspectives and creative thinking become larger.

More generally, lack of agentic diversity is likely to lead to missed market opportunities and heightened business and market risks. Ciklum’s Cetin highlighted three key issues for industries. “First, if everyone is using the same models, you get correlated errors. In regulated industries like payments or insurance, the whole sector experiences the same fraud false negatives at the same time. That’s a systemic risk, not just a vendor risk. Second, in retail, AI recommender and pricing systems converge on the same answers. Retailers using the same stack quietly price toward the same equilibrium, and competitive differentiation compresses without anyone noticing.” One study found that across eight product markets— ranging from electric cars and laptops to running shoes and hotel chains—major AI recommender systems “showed marked favoritism” toward US brands, potentially distorting competition and consumer choice.

The third risk Cetin pinpoints is the loss of insight into business edge cases. For example, the convergence of agentic models in insurance could mean that firms fail to spot novel or unusual patterns of fraud. Or, in consumer product or retail industries, firms may be slow to detect changing consumer patterns or preferences, limiting product experimentation and the cultivation of new customer segments and business models. The upside of agentic diversity is that firms will be more likely to spot different demand signals early, enhancing their commercial sensing and ability to innovate through pricing, marketing and business models.

The implication is clear: As agentic AI is increasingly scaled and diffused across enterprises and their workforces, the risks of growing uniformity and losses from lack of agentic diversity are likely to be magnified across teams, businesses and the wider economy.

Seven Imperatives for Creating Diverse Agentic Teams

 Large enterprises are still generally in the early phase of agentic AI implementation, identifying and piloting use cases while training their human workers in basic AI skills. Yet good practice now in agentic AI development can avoid future problems and reinforce the diversity of emerging agentic teams. These seven imperatives help guide the path forward.

1. Diversify the agentic “tech stack”: The first and most important action to improve agentic diversity is for enterprises to diversify the underlying foundation models, large general-purpose models that act as the brain of an agentic system. Today’s leading examples include Anthropic’s Claude, OpenAI’s GPT, Google’s Gemini, Meta’s Llama, and Mistral’s open models. While the foundation model is the most visible piece, the agentic AI stack also includes other elements that are candidates for diversification, such as the retrieval layer (which pulls the company’s data into the agent system), the orchestration framework that co-ordinates the actions of agents, and the evaluation and guardrails layer that checks outputs before they reach the end user. A practical configuration would be Anthropic’s Claude as the reasoning agent, Google’s Gemini as the evaluator, and OpenAI’s GPT as the generation agent—different labs, different training data, different alignment approaches. The point is structural rather than rankings-based: their errors are less likely to correlate if configured in this way.
2. Enrich agentic training data: More varied training data can also help improve agentic diversity. Researchers have shown that training models using multi-dimensional psychometric datasets—such as those generated by the Big Five Framework—can be used to develop agents that closely mirror humans on personality tests. (The Big-Five personality traits are agreeableness, neuroticism, extraversion, openness, and conscientiousness.) Similarly, AI agents could be trained on datasets such as the World Values Survey to better reflect different cultural values and ways of thinking across different parts of the world. Major AI companies have made a start in this direction, for example by recruiting data labelers from a variety of geographic and cultural milieu, but much more could be achieved in this direction.
3. Fine-tune through small-language models: Enterprises need not rely solely on external datasets and models, however. Most transnational enterprises have vast tracts of internal data that can be used to fine-tune agent models to reflect the composition of their workforce. These include, for example, data from HR systems, employee surveys, and psychometric evaluations of employee personal styles.
4. Train agents by work-shadowing humans: Workers in any organization typically “learn the ropes” of effective teamwork by observing and emulating colleagues. Given their capacity to continually learn on-the-fly, agents can be designed to learn team work styles from human work colleagues in different geographic and cultural contexts. Agents can be trained based on email communications or meeting transcripts to learn principles of effective challenge, negotiation, and consensus. Of course, these benefits will be magnified if the enterprise already has a diversified human workforce.
5. Implement a model portfolio governance policy: Just as cybersecurity and human diversity have become board-level issues, so too must agentic diversity be elevated to the attention of board directors. Ciklum’s Cetin advocates what he calls “a model portfolio governance policy.” He explained: “This is just like a financial portfolio. Boards set a rule that no more than a given percentage of critical agentic decisions can depend on a single model vendor, i.e., a company that builds and operates foundation models such as Anthropic, OpenAI or Google. Concentration risk in foundation models should be governed the way enterprises govern concentration with any other critical supplier. ”
6. Use cultural “red-teaming”: Enterprises can also test their agents through red-teaming, a term borrowed from cyber-security to describe the testing of security defenses by human experts or AI systems. OpenAI, for example, already draws upon a multidisciplinary team of experts to red-team its LLMs for risks such as bias, societal impacts, and cultural sensitivity. Such red-teaming could be expanded to a wider set of societal and cultural factors, and even performed by AI-powered teams in the future.
7. Create agentic talent marketplaces: Major cloud and enterprise software providers already offer platforms for AI agent creation and fine-tuning. In the future we can envisage the emergence of highly liquid agentic talent marketplaces, akin to national and global human talent networks, that enable companies to “recruit” agents and teams that reflect a mix of roles, nationalities, skills, personality types, and cultural backgrounds.

Advances in agentic AI open up a dramatically transformed vista of growth opportunities for business: enhanced worker productivity, leaner and faster processes, enhanced customer and employee experiences, new knowledge and innovation possibilities, to name but a few. Laying the groundwork for authentically diverse agentic models now will pay dividends in times to come.