- AI agents are hot right now, but researchers say there is confusion around the term.
- AI agents do more than just execute tasks. They reason and solve problems.
- Several companies, from OpenAI to Glean, have released agents or agent-related products this year.
“Agents” are one of the buzziest topics in AI, and companies have been rolling out one after another since the start of the year.
But few agree on what they actually do.
In January, OpenAI released Operator, an agent that can browse the web, book travel tickets, and create memes. Enterprise technology companies from Glean to Cohere have unveiled platforms that allow employees to build and deploy agents. And earlier this month, Chinese startup Butterfly Effect drummed up DeepSeek-level buzz by unveiling Manus, an invite-only AI agent that can analyze stocks, scrape data off the web, and even create interactive websites.
Agents are commonly defined as virtual assistants that can complete tasks autonomously. They break down problems, outline plans, and take action without being prompted by a user.
Part of the excitement around them is that they might be the first step to a hands-free world where humans can sit back, relax, and watch AI do all the work. But AI experts who work with agentic systems say the tech is more nuanced.
“ I’ve heard definitions of agents where something is only an agent if it takes actions,” said Douwe Kiela, cofounder and CEO of Contextual AI, which helps companies build and deploy RAG agents. Kiela previously led the team at Meta that developed retrieval augmented generation, or RAG, a technique that enhances the outputs of large language models.
“ Think of Deep Research,” he told Business Insider, in reference to agents developed by OpenAI, Perplexity, and Google that synthesize hundreds of online sources into detailed reports. “That doesn’t necessarily take any actions besides search. Is that an agent? I would say yes, but a lot of people say no, it has to take actions that somehow affect the state or the environment the agent operates in.”
The correct definition of an agent is simply something that “actively reasons,” Kiela said. “ So if it made a mistake, then they can catch that mistake and try again.”
Natural language expert Satya Nitta, CEO of Emergence AI, a company specializing in multi-agent systems, told BI that agents are designed to troubleshoot even as their surroundings constantly evolve.
Agents go “beyond automation, demonstrating contextual reasoning, adapting to unforeseen challenges, and dynamically adjusting plans to succeed in complex environments,” Nitta said.
Another way to think about agents is in relation to large language models, which serve as the basis for popular chatbots like ChatGPT.
“A large language model is a misnomer. It should really be called a neural sequence model, which applies to any sequential data: natural language, programming languages, pixel sequences, and biological sequences (proteins),” Richard Socher, founder and CEO of You.com, an AI-powered search engine for knowledge workers, told BI. An agent, on the other hand, is “a neural sequence model that can take actions for you, learning from patterns of human behavior to automate complex tasks across multiple domains.”
As agents become more ingrained in the workplace, people may start to think of them as team members or authoritative assistants. But Spence Green, CEO of AI-powered translation company LILT, who has years of experience in natural language processing, said, “I think of them as designing workflows.”
However they are defined, and whatever they do, AI companies are betting big on agents to drive returns.
The Information reported this month that OpenAI plans to sell PhD-level agents starting at $20,000 a month and eventually expects 20-25% of its revenue to come from agents.
“If 2024 was the year of LLMs, we believe 2025 will be the year of agentic AI,” Praveen Akkiraju, a managing director at Insight Partners, previously told BI.
