By Steven Rosenbush

SOUTH SAN FRANCISCO, Calif. -- The race to build useful artificial-intelligence agents that can perform complex actions for people is moving on to a new set of challenges.

The large language models at the core of these agents are good enough for many tasks. But there is a growing emphasis on connecting the LLMs inside agents to a plethora of tools that they will need to get their jobs done.

An advanced LLM might fail at complex multiplication, for example, while the cheapest, oldest model can ace the test if it has a calculator tool.

But there is another hurdle: Agents will need permission to access apps, APIs and websites if they are ever going to call an Uber or book a flight, the kind of expectation that has been established over the past year.

Humans type passwords or use facial and fingerprint recognition to sign into their accounts, but AI agents require new methods of authorization to address the intermediary role between humans and the services they want to use, according to Alex Salazar, chief executive of startup Arcade.dev.

The promise of agents

In a recorded presentation shown at Apple's developer conference last June, one of the company's machine-learning and AI leaders shared a hypothetical example in which she asked Siri to tell her when her mom's flight was going to land. Siri would cross-reference flight details that her mom emailed with real-time flight tracking to determine an up-to-date arrival time.

Next, she imagined that she wanted Siri to tell her the details of lunch plans with her mom, which Siri would ascertain by checking her calendar or text messages. And it would figure out how long it would actually take to get to lunch from the airport.

Apple has yet to deliver, but the AI ecosystem is working on the "plumbing" that will make such complex AI agents possible.

The effort got a boost last November when Anthropic, the startup behind the chatbot Claude, introduced an open-source standard known as Model Context Protocol. "Just as USB-C provides a standardized way to connect your devices to various peripherals and accessories, MCP provides a standardized way to connect AI models to different data sources and tools," the user guide says.

"Tool-calling agents are the emerging phase of agent AI development," Salazar told me. His 12-person startup on the outskirts of San Francisco is developing tools to tackle getting agents signed into websites, APIs and apps.

The company, which Salazar co-founded last year with Chief Technology Officer Sam Partee, announced in March that it had raised $12 million of seed funding in a round led by Laude Ventures with participation by Flybridge Ventures, Hanabi Capital and the venture capitalist Andy Rachleff.

Adoption curve

Getting agents all the necessary tools and access is a significant obstacle.

Device manufacturers will likely start integrating AI agents with core applications such as email and calendars, according to Salazar. As agents expand to other services, he said they would work best with companies that have public application programming interfaces, the bits of code that help one application connect to another. Some platforms deliberately limit API access to prevent abuse, and some older systems lack APIs.

But that integration of AI agents and apps via existing authorization protocols could also be the last major challenge.

Salazar says simple, low-risk AI agents work right now, and they will take on increasingly complex, high-risk tasks over the next two years. For example, Arcade is helping a customer, Shortwave, connect its AI email agent to other apps such as knowledge-management tool Notion, according to Shortwave co-founder and Chief Executive Andrew Lee.

AI agents over the next 24 months will increasingly draft communications and plan itineraries for people but still require human confirmation before final execution, Salazar predicts. After that point, he expects that fully autonomous agents will be allowed to operate, beginning with easy, low-risk tasks.

Once the key engineering problems around agents are overcome, Salazar says, the world is poised for a new technological shake-up in the way things get done. The introduction of app stores in 2008 abruptly and broadly changed the norms by which people interact with the world. AI agents could be very close to triggering something just as big.

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

(END) Dow Jones Newswires

May 17, 2025 08:00 ET (12:00 GMT)

Copyright (c) 2025 Dow Jones & Company, Inc.