Alcon, a global supplier of vision-aid products and technology for patients and ophthalmologists, is intent on becoming an Agentic Enterprise. And so, over the past year, various departments within Alcon developed hundreds of AI agents.
But like other organizations on the agentic journey, Alcon soon realized that a collection of specialized AI agents presented its own challenges. “Everyone went and built their own agents — business users and IT users alike,” recalled Sreenivasa Patibandla, the company’s Director of System Integrations and API. “We ended up with 900-plus agents built in silos. It’s a security risk, first and foremost.” And for a regulated medical device company, the hundreds of disparate agents created compliance issues, among other complications.
We ended up with 900-plus agents built in silos. It’s a security risk, first and foremost.
Sreenivasa Patibandla, Director of System Integrations and API, Alcon
These are the problems solved by orchestration – the next great challenge for the Agentic Enterprise. The technology moves organizations beyond the development, deployment, and adoption of individual agents toward harmonious collaboration.
Consider a musical metaphor. It’s one thing for an AI soloist to sing the answer to a customer’s simple question. It’s quite another to create an AI jazz ensemble with the structure and coordination to produce a sonorous group improvisation.
That, in essence, is what agentic AI orchestration is all about: A group of talented, improvising agentic AI soloists work collaboratively, each handing off to the next player without missing a beat. And all the while, an AI bandleader — a superagent — keeps the whole arrangement in the right key, on rhythm, and adhering to the same basic structure.
When AI agents are orchestrated, businesses can leverage their creativity while ensuring they work together productively. At Alcon, the company has begun using a Salesforce orchestration enabler, Agent Fabric, to bring collaborative capabilities, and governance, to all those agents.
Many other Agentic Enterprises have also recognized the need for AI orchestration and are taking steps to implement it. Besides Alcon, these companies include Royal Bank of Canada and the technology consulting firm Diabsolut.
That’s why Salesforce has released a series of products dedicated to solving the orchestration problem. That includes Agent Fabric, which provides an enterprise agent control plane with unified agent management and advanced orchestration capabilities.
Andrew Comstock is GM of Salesforce MuleSoft, a Salesforce company, and the creator of Agent Fabric. Comstock said that today’s AI orchestration challenge is similar to one that IT managers have long faced – getting their services effectively operating in concert and not antagonistically.
Today’s AI agents can be thought of as the next generation of software. But there’s an extra wrinkle. Apps and their APIs are deterministic — the same input always yields the same output.
But AI agents are nondeterministic, reasoning and adapting as they carry out their tasks. “An API does the same thing every time,” Comstock said. “Agents don’t.”
And so, coordinating and governing the activities of these next-generation, semiautonomous, unpredictable AI agents has become the crucial challenge that orchestration is meant to solve.
For Gary Lerhaupt, Salesforce VP of Product Architecture, once an organization’s first agent is working well, challenges will arise as the enterprise attempts to scale that success by simply adding more instructions and capabilities to a single agent, forcing it to handle too many disparate chores.
Lerhaupt referred to these as “monolithic” agents — jacks-of-all-trades that are masters of none. Overstuffing a single agent spreads the AI’s attention too thin, leading to logic errors or hallucinations.
“The ever-expanding monolithic agent eventually degrades,” Lerhaupt explained. “It’s an anti-pattern,” he said, using the industry epithet for a solution that looks attractive on the surface but ultimately creates technical trouble.
Ultimately, Agentic Enterprises must recognize that AI should follow the same organizational patterns as the humans they support. Just as a company wouldn’t hire one person to be the CFO, the head of engineering, and a customer service rep, organizations shouldn’t expect one AI agent to handle every corporate task.
Human specialists know when to reach out to one another via Slack, email, or a quick meeting. Likewise, AI agents must be equally communicative team players. Without this coordination, Lerhaupt said, AI descends into the opposite extreme of the monolith: a “maze” of disparate, single-task agents that customers and users must navigate on their own.
“You don’t want to have to tell your customers, ‘You can find Agent A over here and Agent B over there,’” Lerhaupt said. “That is not a great user experience.”
A company that has deftly avoided the monolith-maze double jeopardy is Royal Bank of Canada (RBC).
Lerhaupt’s team collaborated with RBC’s Peter Herzog, who is the Technical Lead for the bank’s U.S. Wealth Advisory business. Herzog’s technologists support the wealth advisors who work directly with individuals and households to maximize the value of their financial holdings.
Early in 2025, the Herzog team introduced an AI agent that was carefully tailored to help advisors prepare for customer meetings and compile follow-up notes. The agent cut the time previously devoted to those activities in half. RBC advisors were so “blown away” by this efficiency, Herzog said, that his staff quickly pivoted to developing an array of specialized agents to make advisors even more effective and better informed.
“In a wealth-management space that’s highly regulated, that’s no easy task,” Herzog said of his team’s efforts. In particular, the highly personal information of each financial household must be strictly firewalled. This ensures an AI agent can’t disclose sensitive data to anyone outside the household other than the advisor, who holds fiduciary responsibilities. Orchestration that stresses trust and governance was essential.
By late summer 2025 — following a rigorous review by RBC’s regulatory compliance and governance teams — more than a dozen AI agents were in place, including specialists for market insights and customer-portfolio analysis. The superagent that coordinates these specialized tools is the RBC Advisor Assistant. It serves as the single point of entry for users.
“Typically, our wealth advisors are not technically advanced,” Herzog said. “They expect to be able to type in a question and get an answer. We want to provide that experience for them.”
Effective orchestration must bridge the gap between various vendor ecosystems, as organizations increasingly deploy a diverse mix of AI tools from providers like Google, Anthropic, and Microsoft.
To allow for such multivendor translations, two interoperability standards have emerged to ensure smooth handoffs. One is the Agent2Agent protocol (A2A) that Salesforce played a role in creating and was released last year by Google. The second is the Model Context Protocol (MCP) developed by Anthropic. Both protocols are now administered by open-source The Linux Foundation.
According to Diabsolut, a consultancy that helps clients optimize their IT ecosystems in industries that include healthcare, manufacturing, and electric utilities, MCP allows any agent in any domain to access the knowledge it needs without the burden of custom, point-to-point integrations.
Ultimately, Diabsolut predicts: “The companies that will win with agentic AI are the ones building intelligent agents on top of solid platforms, where the AI knows what it’s allowed to touch, what shape the data is in, and what the business rules are before it ever generates a response.”
Herzog at RBC is another proponent of protocol-based multi-vendor coordination. “We want to provide to our customers what feels like a single-agent experience,’’ he said. “But to do that well, you need multi-agent orchestration.”
In the Agentic Enterprise, orchestration is the great enabler. When tools, data, and intelligence play in concert, the technology finally starts working for the user, rather than the user working for the technology. And that’s music to any customer’s ears.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。