“Agentic” was the trendiest AI term of the year. It’s plastered in eye-catching colors all over online adverts and billboards, but what it means in practice can be fuzzy.
Most modern AI “agents” are built on top of large language models (LLMs). An LLM reads a text prompt and predicts what words should come next based on linguistic patterns learned from massive amounts of training data. An agent, on the other hand, uses that LLM to act: it can make decisions, execute practical tasks and achieve a goal.
When using a chatbot, a real person talks to an LLM and copy-and-pastes data between the chat interface and other applications. AI agents have direct access, representing the transition from the limited roles of LLMs as consultants to applying language models to reliably automate work.
For example, Decagon built an autonomous customer service agent that has been adopted by many companies, from Notion to Duolingo. The service agent can access user inputs and knowledge bases — accessing much of the same data a human employee can. The underlying LLM takes in all of this information and decides which actions to take. Their agents carry out that decision by responding to customers, creating customer support tickets, redirecting calls and so on.
Why are agents taking off right now? The biggest reason is that LLMs have improved enough to be trusted with autonomy, and not only in factual accuracy. More critically, in the past few years, they’ve gained the ability to use other tools like internet searches, reasoning loops, API calls and access to software and operating systems.
As training methods and massive investments in computing power have made LLMs far more reliable and capable, the next natural step is to use these more powerful models to do tangible work as AI agents.
That doesn’t mean, however, that the LLM technology is perfect. LLM-powered agents still carry risks of hallucination and error, as chatbots do. Agents could also introduce brand-new problems like the security risks of giving them broader permissions. That’s why most uses of AI agents still involve some human oversight. For example, Anthropic’s coding agent Claude Code has direct access to your codebase, but all edits can and should be reviewed by the human user before being finalized.
Advancements in agentic AI could usher in a new phase in the democratization of AI.
Many people outside of the tech sector use LLMs mainly through a chatbot interface. It’s difficult to set up LLMs to automate tasks — connecting language model APIs to workflows — without AI and programming experience. As more companies offer customizable AI agent products and include agents in their services, LLM automation could become more accessible.