What is an AI Agent?

The Evolving Definition of AI Agents

AI agents are widely anticipated as the next significant advancement in artificial intelligence. However, a precise definition of what constitutes an AI agent remains elusive, leading to ongoing debate among experts.

What Exactly is an AI Agent?

In its most basic form, an AI agent can be understood as AI-powered software designed to perform tasks traditionally handled by human professionals. This could encompass roles like customer service representatives, HR personnel, or IT support staff, extending to virtually any type of assignment. These agents respond to requests and execute them, often integrating across multiple systems and exceeding simple question-answering capabilities.

For instance, Perplexity recently launched an AI agent focused on assisting with holiday shopping, and Google unveiled Project Mariner, its inaugural AI agent, capable of booking travel, purchasing household goods, finding recipes, and more.

Differing Perspectives from Tech Leaders

Despite the apparent simplicity, defining AI agents proves complex, even among leading technology companies. Google views them as task-specific assistants, aiding developers with coding, marketers with color scheme creation, and IT professionals with log data analysis.

Asana envisions agents functioning as additional team members, diligently completing assigned tasks. Conversely, Sierra, a startup founded by former Salesforce co-CEO Bret Taylor and Google veteran Clay Bavor, positions agents as customer experience enhancers, facilitating the resolution of complex problems beyond the scope of conventional chatbots.

Early Stages and Technological Foundations

This lack of a unified definition understandably creates confusion regarding the future capabilities of these systems. However, the core purpose remains consistent: automating task completion with minimal human intervention.

Rudina Seseri, founder and managing partner at Glasswing Ventures, emphasizes that the field is still in its nascent stages, explaining the current lack of consensus. “An agent is an intelligent software system designed to perceive its environment, reason about it, make decisions, and take actions to achieve specific objectives autonomously,” Seseri explained.

These systems leverage a range of AI technologies, including natural language processing, machine learning, and computer vision, to operate effectively in dynamic environments, either independently or in collaboration with human users.

The Potential for Growth and Challenges

Aaron Levie, co-founder and CEO of Box, believes that as AI technology matures, agents will be able to handle increasingly complex tasks on behalf of humans. He identifies several key factors driving this evolution, including improvements in GPU price/performance, model efficiency, model quality, and AI infrastructure.

However, MIT robotics pioneer Rodney Brooks cautions that AI faces more challenging problems than other technologies, and its progress may not mirror the rapid advancements seen in areas like chip development.

“Humans tend to overestimate AI’s competence based on limited observations,” Brooks noted. “They generalize performance to similar situations, often with undue optimism.”

The Difficulty of System Integration

A significant hurdle lies in integrating AI agents with existing systems, particularly legacy systems lacking API access. While improvements are being made, enabling software to navigate multiple systems and address unforeseen issues could prove more difficult than anticipated.

David Cushman, a research leader at HFS Research, suggests that current AI agents are best viewed as assistants, helping humans achieve strategic goals. The true challenge lies in enabling machines to handle contingencies autonomously, a capability that remains distant.

“AI agents represent the next step towards independent and scalable AI operation,” Cushman stated. “The goal is to minimize human intervention and maximize automation.”

The Need for a Dedicated Infrastructure

Jon Turow, a partner at Madrona Ventures, argues that the development of AI agents requires a dedicated infrastructure—a tech stack specifically designed for their creation and deployment. He highlights the need for a robust platform that ensures scale, performance, and reliability.

“Building infrastructure to support AI agents and the applications that rely on them is crucial,” Turow wrote in a recent blog post.

The Role of Multiple Models

Fred Havemeyer, head of U.S. AI and software research at Macquarie US Equity Research, believes that effective AI agents will likely rely on multiple models, rather than a single large language model (LLM). This approach aligns with the idea that agents perform a variety of distinct tasks.

“No single publicly available LLM can currently handle agentic tasks effectively,” Havemeyer explained. “Multi-step reasoning, essential for a truly agentic future, remains a challenge.”

He envisions agents as collections of specialized models, with a routing layer intelligently directing requests to the most appropriate resource, functioning as an automated supervisor.

Looking Ahead

Ultimately, the industry is striving towards fully autonomous agents capable of independently pursuing abstract goals and devising the necessary steps to achieve them. However, significant advancements and breakthroughs are still required. While current progress is promising, we are not yet at the point where AI agents can fully realize their envisioned potential.

This story was originally published July 13, 2024, and was updated to include new agents from Perplexity and Google.

Karyne Levy contributed to this story.