Agentic AI

Agentic AI refers to AI systems that can pursue goals with a degree of independence — planning steps, making decisions, and taking actions through tools over time — rather than simply responding to one request at a time.

Agentic AI describes a shift in what we ask AI to do. Most AI you've used is reactive: you make a request, it responds, and the interaction ends until you ask again. Agentic AI is the move toward systems that take a goal and run with it — planning a sequence of steps, deciding what to do next, taking actions in the world through tools, observing the results, and continuing until the objective is met. The word "agentic" captures the quality of acting with purpose and a degree of independence, rather than just producing one answer on demand. It's less about a single product and more about a property a system can have: the capacity to drive toward an outcome over multiple steps instead of waiting to be prompted at each one.

The intuition is the difference between an employee handed a detailed checklist and one given an outcome and trusted to work out how to get there. The checklist-follower does exactly what's written and stops; hand them something the checklist didn't anticipate and they're stuck. The trusted employee takes "sort out the venue for the team offsite," and figures out the steps themselves — comparing options, sending enquiries, adapting when a first choice falls through. Agentic AI aims for that second mode. In practice these systems are usually built around a capable language model doing the planning and decision-making, connected to tools — web search, software, databases — that let it act rather than merely talk, all running in a loop that alternates between reasoning about the next move and carrying it out.

This is one of the most active and hyped frontiers in AI, so it's worth being level-headed about it. Trading a fixed script for a system that finds its own route costs you something real: a script's great virtue is that you know in advance exactly what it will and won't do, and that guarantee is precisely what you hand back, which is why serious agentic systems rebuild it deliberately as explicit limits and approval points. It also helps to remember that "agentic" names a property a system can have, not a product you buy — and it arrives by degrees rather than all at once. So the honest question to ask of any tool advertised as agentic is a plain one: how much of the goal will it actually carry before it needs you to prompt it again?