Agentic AI
AI that does not just answer but acts: it takes a goal, plans the steps, uses tools, and carries out a task across many turns with limited supervision, rather than producing a single reply and stopping.
What It Is
Agentic AI is the difference between a system that answers you and a system that acts for you. A chatbot takes your question and returns a reply, and there it stops. An agentic system takes a goal, works out the steps to reach it, uses tools to carry those steps out, checks its own progress, and keeps going across many turns until the goal is met or it hits a wall. The word “agentic” points at that quality of acting on its own behalf inside a task, not at any single model.
This edition’s news is what makes the term worth pinning down now. When Microsoft moved its Project Perception security platform into public preview and described it, in its own words, as agents that act, not just alert, it was drawing exactly this line. One system watches for a threat and tells a human. Another watches, judges the real risk, and takes the corrective step itself, pausing only for a human sign-off on the consequential moves. The first is an assistant. The second is an AI agent, and a coordinated set of them working toward a goal is what “agentic AI” names.
How It Actually Works
Under the surface, an agentic system runs a loop rather than a single pass. It reads the goal, decides on a next action, takes that action, reads the result, and decides again. The actions are made possible by tool use: the model is given a set of functions it can call, such as searching a database, sending an email, editing a file, or running code, and it chooses which to call and with what inputs. Each loop feeds the outcome back in, so the system can correct course when a step fails.
The hard part is not the single step. It is sustaining a coherent line of action across dozens of steps and a long stretch of elapsed time without losing the thread, which is the problem a long-horizon agent is built to solve. That is why the shipping examples still keep a human on the consequential moves. The loop is reliable enough to do the work and not yet reliable enough to be left entirely alone.
Why It Matters Right Now
Two things converged this week to move agentic AI from a demo word to a deployment word. First, the models underneath got dramatically cheaper, with frontier-class intelligence now priced near $2 per million input tokens. An agent that runs a long loop calls the model many times, so the cost of the loop was the thing standing between “interesting” and “worth shipping,” and that cost just fell. Second, real products crossed the line from advising to doing inside enterprise software people already run. When the capability arrives in a tool an operator already uses, the question stops being theoretical.
The Cost / Tradeoff
The trade is autonomy against oversight, and it is not free in either direction. An agent that acts saves you the steps, but every step it takes on its own is a step you did not watch. A single wrong action, an email sent to the wrong list, a file overwritten, a payment made, can cost more than all the minutes the agent saved. There is also a real money cost: because an agentic loop calls the model repeatedly, its bill scales with how long and how hard it works, not with a single request. Cheaper tokens ease that, but a long-running agent is still a metered thing, and the meter runs while you are not looking.
How TWO Uses It
TWO’s rule for agentic AI is the leash, and it is deliberately short at the start. Scott’s practice is to hand an agent one real recurring chore he already trusts himself to check, keep it read-only until he has watched it work for a week, and judge it on three questions: did it touch only what it was asked to, did it stop and ask before anything consequential, and would he have to redo its work. Only when those answers hold does the leash get longer. The failure mode TWO warns against is the opposite instinct, handing an agent broad reach on day one because the demo looked clean, then discovering the cost of the one action nobody watched.
The wider frame is that an agent acting in your name makes you a steward, not a spectator. TWO counts these systems the way it counts an agentic workforce: as delegated capacity you are accountable for, not as help that absorbs the accountability for you. Delegating the work never delegates the responsibility, which is why disclosure and human sign-off are features, not friction.
What to Watch Next
The signal that agentic AI is maturing is not a smarter demo. It is the shrinking of the human-in-the-loop step. Watch for how much a shipping product still pauses for a person on consequential moves, and watch the regulation moving alongside it, such as the EU’s new duty to disclose that an AI is acting at all. When the pause gets shorter and the disclosure gets stronger at the same time, the technology and its guardrails are maturing together, which is the healthy version. When one races ahead of the other, that gap is where the trouble collects.