The concept of AI agents is gaining traction, with tech giants like Google and startups like Perplexity releasing their own versions of these autonomous software systems. However, despite the excitement, there's a glaring issue: nobody can agree on what exactly constitutes an AI agent.

At its core, an AI agent is a software system that performs tasks on behalf of humans, often crossing multiple systems and going beyond simple question-answering. For instance, Perplexity's AI agent helps with holiday shopping, while Google's Project Mariner assists with tasks like finding flights and hotels. But this lack of clarity has led to differing views from tech giants and experts.

Google sees AI agents as task-based assistants, depending on the job at hand. Asana, on the other hand, views them as extra employees that take care of assigned tasks. Sierra, a startup founded by former Salesforce co-CEO Bret Taylor and Google vet Clay Bavor, sees agents as customer experience tools that help solve complex problems.

Rudina Seseri, founder and managing partner at Glasswing Ventures, attributes the lack of agreement to the early days of AI agent development. "There is no single definition of what an 'AI agent' is," she said. "However, the most frequent view is that 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."

Aaron Levie, co-founder and CEO at Box, is optimistic about the potential of AI agents, citing a self-reinforcing flywheel that will drive their evolution. However, MIT robotics pioneer Rodney Brooks cautions that AI has to deal with tougher problems than other technologies, and growth may not happen as rapidly as expected.

One of the significant challenges AI agents face is accessing multiple systems while solving problems along the way. This is complicated by the fact that some legacy systems lack basic API access. David Cushman, a research leader at HFS Research, sees the current crop of bots as assistants that help humans complete tasks, but notes that true automation requires letting the AI agent take over and apply its own decision-making.

Jon Turow, a partner at Madrona Ventures, believes that building an AI agent infrastructure is necessary to support the growth of these autonomous systems. This tech stack would be designed specifically for creating AI agents, which would require a collection of different models rather than a single large language model.

Fred Havemeyer, head of U.S. AI and software research at Macquarie US Equity Research, agrees that multiple models will be necessary to make agents work. He envisions an "interesting [automated] supervisor, delegating kind of role" that would route requests to the most effective agent and model.

Ultimately, the industry is working towards a future where AI agents operate independently, taking abstract goals and reasoning out individual steps on their own. However, we're still in a period of transition, and it's unclear when we'll reach this end state. While the current developments are promising, we need further advances and breakthroughs for AI agents to operate as envisioned.