AI agents: How tool calling enabled a system of capabilities orchestration in enterprises

In 2022, Verta’s founder Manasi Vartak had said during a media interview, “We still haven’t seen the enormous power (and the challenges) of building products that are fully AI-enabled.”

Verta was acquired by Cloudera in 2024, and today as part of Cloudera’s AI platform, it helps teams build and deploy AI-enabled products and applications faster and more safely than ever before. 

Its mission changed to one with a focus on governance and safety, when Manasi now as Cloudera’s Chief AI Architect discovered customers had challenges operationalising research models – integrating them into user-facing features (like an AI assistant) or business processes and then running them at scale.

Manasi had shared her quote about AI-enabled products in 2022, citing AI assistants as one example. But back then, had she ever imagined a technology like agentic AI would come along? She answered, “There have been various efforts, at least in the research community, to make computing more autonomous. We call them agents now.”

She also shared a revealing insight when she said that what unlocked agents, as we call them right now, is reasoning or large language models (LLMs), tool calling functions, and modes of collaboration.

The power of tool calling

In summary, tool calling is the mechanism by which LLMs extend their capabilities by interacting with external tools and APIs; it is considered a foundational advancement for agentic AI to accomplish tasks that training data alone, cannot.

Manasi also weighed in, “Tool calling has been one of the really big advances (recently). If you ask an AI assistant to find you a flight from San Francisco to New York at this time of day, LLMs can’t do that by themselves without access to flight schedule data.”

What LLMs are able to do however, is recognise the question and that it needs to call an API for the data required to accomplish its task.

“So, I think these have been one of the fundamental building blocks developed in the last year that is working really well right now – tool calling and MCP or Model Context Protocol, which standardises how tools are called.”

Evolving permissioning models

As agentic AI becomes more widespread, governance and safety become critical. Enterprises need to ensure that agents access data appropriately and do not violate privacy or security policies. The challenge is exacerbated by the increasing number and variety of agents, each potentially accessing different types of data.

Traditional permissioning systems are designed for humans, but Manasi pointed out that with agents acting on behalf of users, these models need to evolve.

“If you have a bot and it can go and crawl Internet links, suddenly everything is accessible. So, we have to consider whether our permissioning model, in some cases, needs to change to accommodate agents and automation.”

The Chief AI architect sees the growing need for governance and safety, as enterprises begin deploying more and more AI agents. So, the question is, how do I make sure that agents are accessing data in a way that is allowed within my system?

“This is where Cloudera has a key role to play because it involves data governance and we’ve historically had a very strong presence there.”

Emphasising governance in the AI journey

To integrate machine learning models (ML) and large language models into production, organisations need to address several challenges including evaluation, monitoring and adapting to rapid improvements in model capabilities. Also crucial, is close collaboration between data scientists, software engineers, and product teams.

According to Manasi, the process is similar for both ML and LLMs, but LLMs bring unique challenges due to their complexity and rapid evolution. Models have to perform as expected in real-world scenarios, and organisations need to evaluate models on specific tasks, monitor their performance, and be prepared for frequent updates and improvements.

One of the ways to fulfil this is evaluation-driven development, or AI evals during  Integration of LLMs into production. These AI evals are like test cases for software that check if the model can perform required tasks. They play a crucial role by providing structured, test-like assessments that ensure the model meets the required standards and evolves alongside the product.

Manasi explained, “If we take the travel example, booking a flight from A to B with no constraints on say, day and time, may be a basic eval – so now, can the agent go and do this?

“When it does, then you add more complex requests like no layovers, travel at a certain time, and so on.”

AI in the next five years

“So, you don’t need to go click on five things and see what the browser says. You can just tell this assistant, ‘Hey, book me a flight’ and you are never going to see travel websites like Kayak.com or whatever it is you usually use,” Manasi said.

What if the current vision for AI assistants is to become the primary gateway to the Internet, replacing traditional browser-based workflows? Companies like Perplexity are exploring this future, aiming to provide users with personal agents that handle complex tasks seamlessly, much like an executive assistant.

What does this capability mean for enterprises that are currently exploring the viability of dozens, if not hundreds, of these AI agents as part of their workforce?

A Linkedin post Manasi shared perfectly sums up what enterprises need to start thinking about now: AI is no longer a single product feature. Instead, it is becoming a system of capabilities that enterprises must orchestrate responsibly, securely, and economically. Above all, we are very early in AI.

“If GenAI were a cat, it would be a (very expensive) kitten,” she concluded.

(This journalist was Cloudera’s guest to its flagship customer event, EVOLVE25, in New York).