IBM and NVIDIA push AI agents into the enterprise fast lane

AI agents—autonomous systems powered by large language models (LLMs) that can take actions, call tools and interact with software—are no longer simply answering questions. They’re beginning to function more like digital coworkers.

NVIDIA recently introduced AI-Q and AgentIQ, a new open-source blueprint and toolkit aimed at helping businesses deploy multiple agents simultaneously to tackle complex problems. These tools are designed to enable agents to reason, plan and work together across software platforms. The move signals NVIDIA’s deeper foray into agents, joining IBM, which has already been developing agentic systems for business use through its IBM watsonx Orchestrate platform.

IBM watsonx Orchestrate allows prebuilt agents to plug into more than 1,500 enterprise tools, streamlining tasks across departments without requiring code or major software overhauls. As such, it’s part of a broader trend in enterprise AI emphasizing scalable, integrated systems capable of handling real-world operations.

“It’s a window to bring automation to every single corner of enterprise,” says Maryam Ashoori, a Senior Director of Product Management for IBM watsonx.ai, in an interview with IBM Think. “You can connect LLMs and gen AI to all your existing workflows.”

At the center of these developments is a concept known as multi-agent orchestration—an approach that draws on principles from both software engineering and human collaboration. Unlike earlier AI assistants, which were designed to complete single tasks, orchestrated systems coordinate multiple agents, each responsible for a specific function. One agent might retrieve data, another might analyze it, while a third might summarize and route the findings.

Ashoori describes agent systems as a way to connect LLMs with existing enterprise workflows, where each part of a task can be handled by a specialized agent. “Because of the AI skills gap and the complexity of today’s tech stack, enterprises are looking for ready-to-go solutions. That’s what these agents provide,” Ashoori says. “You can have a prebuilt agent that talks to an API, connects to SAP or Workday, and chains tasks together to get work done. You don’t need to be a developer—you just drag, drop and configure it to your needs.”

Multiagent systems introduce new challenges around architecture, reliability and observability. That’s where infrastructure like NVIDIA’s AI-Q and IBM watsonx Orchestrate come into play. Both aim to provide enterprises with a scaffold that can scale and adapt to changing business needs without starting from scratch.

IBM has already introduced orchestrated AI agents in a range of industries. Avid Solutions says it cut project errors by 10 percent using similar systems. Dun & Bradstreet, a financial services firm, used AI agents to reduce procurement task time by up to 20 percent.

In some cases, these agents are used to automate back-office functions, like processing invoices or routing internal requests. In others, they’re applied to frontline operations, such as handling customer service interactions or scanning documents for regulatory compliance.

Still, the technology is not fully autonomous. Human oversight remains a key element, particularly in sensitive or regulated environments. “To this date, I have not seen one single implementation of agents that is capable of taking the right actions on a data set with no human in the loop,” Ashoori says. “The scale of the potential undesired consequences is high.”

This concern has led companies to build “human-in-the-loop” systems, where AI agents propose actions but require human approval before executing them. That structure can help mitigate risk while also improving user trust—an area Ashoori has studied extensively.

With AI-Q and AgentIQ, NVIDIA is offering a more developer-focused alternative, with the aim of helping businesses design multi-agent systems that can be adapted to various operational needs. AI-Q is a reference architecture incorporating NVIDIA’s accelerated computing capabilities and storage partnerships, along with software tools like NeMo Retriever, NIM microservices and the Llama Nemotron reasoning models. AgentIQ, released as an open-source tool on GitHub, provides observability and performance tuning for AI agents. It also enables profiling and traceability—features increasingly sought out by companies looking for transparency around automated decision-making.

NVIDIA says AgentIQ integrates with various platforms and frameworks, including Salesforce’s Agentforce, Atlassian’s Rovo, Microsoft’s Azure AI Agent Service and ServiceNow. It also works with LangGraph, CrewAI and Letta, allowing developers to work in their preferred environments.

Trust remains one of the most significant barriers to AI adoption in enterprise environments. A few years ago, Ashoori co-authored a paper titled “In AI We Trust?” which explored the factors that influence humans' ability to trust AI decision-making.

“You might be fine with an AI suggesting where to order dinner,” Ashoori says. “But when it comes to hiring decisions, legal reviews or financial risk assessments, people want to understand how the decision was made—and have the power to change it.”

AgentIQ addresses this by allowing IT and compliance teams to monitor agent activity in real time, audit decisions and fine-tune performance. For many industries, including financial services and healthcare, that level of control is non-negotiable.

Despite the growing adoption of agentic systems, technical challenges remain. Ashoori points to reasoning and planning as areas where today’s LLMs still fall short of humans. “If I want to pitch one breakthrough that still needs to happen, it’s around reasoning and decision-making,” she says. “Today’s LLMs aren’t deciders. They can act and retrieve, but their planning and logic are still limited.”

AI agents are already being deployed across industries to automate critical tasks and improve operational efficiency. Visa is one of several companies that are already using AgentIQ to automate key processes. The company has applied NVIDIA’s profiling tools to phishing email analysis, allowing it to rapidly identify and mitigate threats while reducing manual review.

In these scenarios, the goal isn’t to replace employees, but to expand their capacity, Ashoori says. “The value isn’t in removing people,” she says. “It’s in letting people focus on the decisions that matter, while agents handle the rest.”

Though IBM and NVIDIA differ in approach—the one focused on user-friendly enterprise integrations, the other on open developer tooling—they’re both betting on a future where networks of AI agents can handle increasingly complex tasks.

Ashoori notes that enterprises vary widely in how they implement AI, and says flexibility is crucial when designing agentic systems that fit within existing workflows.

To scale AI beyond isolated use cases, many companies are turning to multi-agent orchestration, whether through low-code platforms or custom-built systems, Ashoori says. However, the long-promised transformation of enterprise AI may depend less on single-model performance and more on how well teams of agents, and their human counterparts, work together.

As the technology matures, observers say the focus will shift from capability to coordination: making sure the right agent is doing the right task at the right time, under the right supervision.

"This is still early," Ashoori says. "But we’re starting to see how agentic systems can help enterprises work smarter, not just faster."