For Better AI Results, Split The Work Across Agents

In this blog we discuss:

• Three Agents Are Better Than One

• The Valuable Role of the ‘AI Coordinator’

• How to Use This Research in Your Organization

• Explainability Is The New AI Advantage

When organizations ask an AI agent to make the best possible recommendation and follow strict rules, something usually gives: the rules bend or the quality of the answer drops. It begs the question: how do you give AI the freedom to find the best answer while keeping firm guardrails in place? 

The Workday AI Research team dug into this question in the pursuit of trustworthy, useful, and efficient enterprise AI. The answer lies in changing AI’s role from personal assistant to orchestrator, according to our new study. Agents produced higher quality and more compliant answers—and boosted accuracy by 5.8%—when they orchestrated other agents.

Three Agents Are Better Than One

Most teams use AI as an assistant that answers a question.

That's because AI agents are often asked to do several things at once: deliver relevance, provide variety, and follow rules. And when those goals compete, trade-offs are inevitable. 

Workday’s study split the work across two specialized AI agents, instead of asking one agent to balance every objective at once. While one agent focused on compliance, only surfacing options that follow every rule, a second agent searched more broadly for new possibilities—including options that would temporarily break the rules. That wider search uncovered promising recommendations the rule-bound agent would never reach on its own. And as they went along, the two agents shared what they learned with each other.

Above them sat another agent: an AI coordinator. It periodically checked how the search was progressing and whether the rules were being met. Then it used that information to adjust where the two agents should focus their effort. If answers started violating the rules, for example, the coordinator shifted attention back to finding compliant options.

A separate mechanism gradually tightened the rules as the search unfolded. That meant the agents could explore freely early on, but their final answers still had to satisfy every constraint.

Think of our experiment like a small team working in real time: an AI innovator exploring possibilities, an AI operator making sure those options actually work, and an AI manager deciding where to focus next.

In the study, this dual-agent approach outperformed a single agent across the board—on accuracy, quality, and compliance. In fact, every final AI answer followed all three of the study's rules—100% of the time. Answer quality also improved by roughly 4-6% overall, with the biggest gains (about 5.6%) showing up on the largest, most varied catalogs.

Moving from a single agent to two specialized agents—and then adding an AI coordinator—improves quality and accuracy, while every setup keeps the rules 100% of the time. (Quality and accuracy scores are on a 0–1 scale; higher is better.)

The Valuable Role of the ‘AI Coordinator’

The researchers then asked a second question: how much of the improvement comes from the AI coordinator, and how much from using two agents instead of one? 

The biggest difference, they found, came from splitting the work between two agents. When they swapped the coordinator for a simple fixed rule that split the work evenly, quality dropped by only about 1%. The coordinator's real value showed up in how it adapted—leaning toward options that followed the rules when compliance was at risk, pushing for broader exploration when progress stalled, and explaining each decision in plain language.

The AI coordinator’s adaptability and explainability came at a cost, though: the initial test ran roughly 94% slower than the single-agent version. 

Still, adaptability trumps the added time. When teams can see why an AI agent made a choice, they're more likely to trust it—and better positioned to give it feedback to continuously improve its answers over time.

How to Use This Research in Your Organization 

While our study used Amazon shopping data, the difference between 100% and "good enough" grows even more consequential when AI is helping organizations make real workforce decisions, such as assigning learning paths, or recommending benefit options.

When recommending a benefits package to an employee, the system has to weigh what best fits the employee's family situation, offer enough variety to give them real choice, reflect the employer's cost and coverage strategy, and stay strictly within eligibility rules and regulatory requirements. No single objective outranks the others—the recommendation is only useful if it satisfies all of them at once.

For organizations using AI to balance multiple priorities and rules, the research points to four best practices:

• Make critical requirements hard rules, not soft preferences agents can quietly override.

• Divide complex work among specialized agents instead of expecting one model to juggle everything.

• Treat rule strictness as a deliberate, human-set choice—tighter for high-stakes decisions, looser where discovery matters more. In the study, this setting affected outcome quality more than any other.

• Add an AI coordinator selectively, in situations where its ability to adapt as conditions change—and explain its choices in plain language—justifies the added time.

Explainability Is The New AI Advantage

The research argues that AI produces better results when organizations use it as an orchestrator—one that coordinates specialized agents, decides where to direct effort, and reserves final judgment for human decision makers.

That finding raises questions worth asking of any AI system. What goals has it been given? Which rules are absolute, and who sets the trade-offs between them? Can people see why it made the choices it did? And what happens when following every rule makes a good answer impossible?

Explainability is what makes those questions answerable. When an AI coordinator can lay out its reasoning in plain language, organizations can understand its goals, challenge its trade-offs, and catch its blind spots. And as AI takes on more coordinating roles, that explainability is what builds trust—and keeps the people managing these systems accountable for the outcomes.

AI is already helping organizations make better decisions. But the differentiator is no longer which model they choose. It's the goals they set, the limits they draw, and the clarity they have into how AI is making decisions.

A note on the research: The study tested DualAgent-Rec on three product categories from a public Amazon shopping dataset, using rules the researchers defined. It did not test Workday products, customer data, or workplace decisions—the enterprise parallels in this piece are illustrative.

Looking for more insights on how to make AI trustworthy, useful, and efficient for the enterprise?  Visit workday.com/ai-research for a complete list of Workday AI research papers.