What is an AI Agent? Demystifying the Buzz

The Buzz Around AI Agents: A Definition in Flux

Investment in AI agents within Silicon Valley is demonstrably strong. Sam Altman, CEO of OpenAI, has stated that these agents will be integrated into the workforce in the current year. Similarly, Satya Nadella, leading Microsoft, anticipates agents will assume responsibilities currently handled by knowledge workers. Marc Benioff, Salesforce’s CEO, has articulated a goal for Salesforce to become the foremost provider of digital labor globally, leveraging the company’s suite of “agentic” services.

The Core Issue: Defining the AI Agent

Despite the widespread enthusiasm, a consensus regarding the precise definition of an AI agent remains elusive.

The Rise of the "Agent" Buzzword

Over the past several years, the technology sector has repeatedly heralded AI “agents” as a transformative force. Much like AI chatbots, such as OpenAI’s ChatGPT, revolutionized information access, agents are predicted to fundamentally reshape work processes, according to prominent CEOs like Altman and Nadella.

However, the realization of this potential is contingent upon a clear understanding of what constitutes an “agent,” a definition that proves surprisingly difficult to establish. Terms like “agent” and “agentic” are experiencing semantic dilution, mirroring the fate of other AI-related terminology – including “multimodal,” “AGI,” and “AI” itself – and risk becoming devoid of concrete meaning.

Confusion Among Tech Giants

This ambiguity presents a challenge for companies like OpenAI, Microsoft, Salesforce, Amazon, and Google, all of whom are developing extensive product lines centered around agents. An agent developed by Amazon may operate fundamentally differently from one created by Google, leading to confusion and potential frustration for customers.

Ryan Salva, a senior director of product at Google and former leader at GitHub Copilot, has expressed his frustration with the overuse of the term “agents.”

Industry Concerns About Terminology

“The term ‘agent’ is overused within our industry to the point of being almost meaningless,” Salva shared with TechCrunch. “It is a significant pet peeve of mine.”

The difficulty in defining AI agents is not a recent development. Ron Miller, a former TechCrunch reporter, previously highlighted the issue, noting that each company approaches agent technology with a unique methodology.

Contradictory Definitions from Leading Companies

The situation has become more complex recently. OpenAI recently published a definition of agents as “automated systems capable of independently completing tasks for users.” Concurrently, the company released developer documentation describing agents as “LLMs equipped with instructions and tools.”

Leher Pathak, OpenAI’s API product marketing lead, subsequently suggested that the terms “assistants” and “agents” are interchangeable, further complicating the landscape.

Distinctions and Overlaps

Microsoft differentiates between agents and AI assistants, characterizing agents as specialized “new apps” for an AI-driven world, while assistants provide more general support, such as email drafting. Anthropic acknowledges the multiple interpretations of “agent,” encompassing both fully autonomous systems and prescriptive, workflow-based implementations.

Salesforce employs perhaps the broadest definition, defining agents as “systems that can understand and respond to customer inquiries without human intervention.” Their website categorizes agents into six distinct types, ranging from “simple reflex agents” to “utility-based agents.”

The Root of the Disagreement

The current state of affairs stems from the inherent nebulousness of agents – and AI in general – and their continuous evolution. OpenAI, Google, and Perplexity have only recently launched their initial agent offerings – Operator, Project Mariner, and a shopping agent, respectively – and their capabilities vary considerably.

Rich Villars, GVP of worldwide research at IDC, points out that tech companies have historically prioritized practical outcomes over strict adherence to technical definitions.

The Role of Marketing

“Companies are more concerned with what they aim to achieve technically, particularly in rapidly evolving markets,” Villars explained to TechCrunch.

Andrew Ng, founder of AI learning platform DeepLearning.ai, attributes a significant portion of the issue to marketing efforts.

“The concepts of AI ‘agents’ and ‘agentic’ workflows once held a specific technical meaning,” Ng stated in a recent interview, “but approximately a year ago, marketers and several large companies adopted these terms.”

Opportunity and Challenges

Jim Rowan, head of AI for Deloitte, views the lack of a unified definition as both an opportunity and a challenge. The ambiguity allows for customization, but it also risks “misaligned expectations” and difficulties in evaluating the value and ROI of agentic projects.

“Without a standardized definition, even within a single organization, benchmarking performance and ensuring consistent results becomes difficult,” Rowan said. “This can lead to differing interpretations of what AI agents should deliver, potentially complicating project goals and outcomes. While flexibility can foster creative solutions, a more standardized understanding would help enterprises navigate the AI agent landscape and maximize their investments.”

A Definition Unlikely to Emerge

Given the historical trend with the term “AI,” it appears unlikely that the industry will converge on a single definition of “agent” in the foreseeable future.