What is an AI agent, exactly?

AI agents are considered the next big development in artificial intelligence, but there’s no universally agreed-upon definition of what they are. Broadly, an AI agent is software powered by AI that performs tasks previously done by humans, such as customer service, HR duties, or IT support. These agents can handle a wide range of tasks, often crossing multiple systems. For example, Perplexity recently launched an AI agent for holiday shopping, and Google introduced Project Mariner, which assists with finding flights, shopping, and recipes.

Despite their potential, defining AI agents remains a challenge, with varying interpretations across the tech industry. Google views them as task-oriented assistants tailored to specific jobs, such as coding help or IT troubleshooting. Asana treats them as virtual employees, while startups like Sierra see them as advanced customer experience tools capable of solving complex problems. This inconsistency in definition highlights the early stage of AI agent development.

Rudina Seseri, managing partner at Glasswing Ventures, describes AI agents as intelligent software systems that perceive their environment, reason, make decisions, and take autonomous actions to achieve specific objectives. They rely on AI technologies like natural language processing, machine learning, and computer vision. Aaron Levie, CEO of Box, believes that advancements in areas like GPU performance, model efficiency, and AI frameworks will significantly enhance AI agents’ capabilities over time.

However, MIT robotics expert Rodney Brooks cautions against overestimating AI’s potential, emphasizing the difficulty of handling complex tasks across multiple systems, especially with legacy systems lacking API access. Current AI agents are more like assistants that follow user-defined guidelines but struggle with fully automated, contingency-aware operations.

David Cushman of HFS Research notes that true automation, where AI operates independently at scale, is still a work in progress. Achieving this will require new infrastructure and tech stacks specifically designed for AI agents, as Jon Turow of Madrona Ventures explains. He envisions a future where developers focus on product differentiation while the underlying platforms handle scalability and reliability seamlessly.

Fred Havemeyer of Macquarie US Equity Research predicts that successful AI agents will combine multiple specialized models with a routing system to delegate tasks efficiently. He believes fully autonomous agents capable of abstract reasoning and independent decision-making are the ultimate goal but acknowledges the industry isn’t there yet.

In summary, while AI agents hold great promise, the technology is still evolving. Significant breakthroughs are needed before these agents can achieve their full potential as envisioned by experts. For now, they represent an exciting but incomplete step toward true AI-driven autonomy.