AI agents are learning how to collaborate. Companies need to work with them

Companies should start planning for the next stage of artificial intelligence: the orchestration of multiple agents across their businesses.

Most companies are still figuring out how to deploy even one AI-powered agent that can perform a task autonomously or in coordination with humans. But developers are creating protocols to harness these agents into teams that handle everything from customer service and coding to supply chain, logistics, finance, marketing and business strategy.

Given the pace of innovation and the time it takes for organizations to adapt, companies will do themselves a favor by getting ready now for multiagent systems increasingly available later this year.

Accenture’s chief AI officer, Lan Guan, says only 10% to 15% of her clients currently use multiagent systems, but she expects that percentage to exceed 30% within 18 to 24 months.

The professional services company has created a 15-agent system used for marketing, for example, comprising three “super agents" that are responsible for coordinating 12 agents trained for specific tasks.

It can plan a marketing campaign around a topic such as “2025 trends," conducting research, identifying similar past campaigns and addressing questions like a human would, according to Guan.

All told, Accenture has more than 50 multiagent systems today for a range of industries and markets, and expects that number to hit more than 100 by the end of the year. The firm said customers such as carmaker BMW, consumer-brands company Unilever and sports giant ESPN are currently adopting these systems.

Accenture last month introduced Trusted Agent Huddle, which it said allows agent-to-agent interoperability with partners such as technology companies Amazon Web Services, Google Cloud, Meta, Microsoft, Nvidia, Oracle, Salesforce, SAP and ServiceNow.

Multiagent capabilities are about to become more widely available. Salesforce and Google announced at the Google Next conference in April that they were working on a protocol called A2A, or Agent-to-Agent. The protocol, which allows agents within Salesforce’s Agentforce ecosystem to interact with each other as well as external agents, focuses on areas such as authentication, identification and message passing, according to Gary Lerhaupt, vice president of product architecture for Agentforce. Work is under way with partners to develop prototype multiagents using A2A, he said.

Keyway, a commercial real-estate tech startup based in New York, provides a glimpse into how the concept works in practice, according to co-founder and Chief Executive Matias Recchia. It offers asset managers and property managers a multiagent platform that uses coordinated interactions to address questions such as how to price a rental property or target amenities and incentives.

The company has raised $45 million from investors including Canvas Ventures, Camber Creek and Thomvest.

While Keyway’s agents are specialized and can trigger each other through structured workflows, the company said, they still operate in a controlled, predefined sequence that requires human oversight such as setting prompts, reviewing outputs and supervising decisions.

A true multiagent system, Recchia said, involves agents that dynamically reason, negotiate or collaborate in real time without requiring human-defined workflows, explicit prompts or manual coordination. In other words, the agents take initiative, adapt to new information and interact fluidly with other agents without waiting for human instruction.

How companies can prepare for multiagents

Companies can start to prepare for multiagents systems simply by creating standard, stand-alone agents. Once the proper protocols are ready, companies can orchestrate these agents into tackling complex, collaborative systems.

Principle Financial Group has embedded individual AI agents across domains including software engineering co-pilots, claims summarization and post-call analytics, according to Chief Information Officer Kathy Kay. They largely operate within defined scopes, but the investment management and insurance company is actively building the technical foundation to support agent-to-agent collaboration, Kay said.

That means developing data pipelines and governance models. Workflows will also have to evolve to accommodate real-time collaboration between humans and intelligent, adaptive AI systems, she said.

Kay foresees strong potential for the use of multiagent systems in retirement services such as rollover optimization. In asset management, she expects multiagents to analyze unstructured market data, generate investment narratives and align findings across portfolios. Other uses include earnings call preparation, contact center intelligence and claims adjudication.

“These are not isolated functions," Kay said. “They are systems of tasks that, when connected through intelligent agents, can drive faster insights and better outcomes across the enterprise."

Write to Steven Rosenbush at steven.rosenbush@wsj.com