Agents
Agentic AI
The umbrella term for the shift from answering to acting: systems that execute workflows independently instead of waiting for every individual instruction. The promise is relief from whole chains of tasks; the price is a control architecture: logs, limits, escalation. Agentic AI is less a product category than an operating model.
An example
The difference from the automation that has existed for thirty years: a classic workflow is a flowchart: on condition A go left, on B go right, every path drawn in advance. An agentic system is given a goal and tools and finds the path at runtime. It therefore copes with cases nobody anticipated, and in unanticipated cases it does things nobody anticipated.
A closer look
The difference from chat is not intelligence but effect: the system is allowed to write, send, book, change. That makes permissions, logs, and reversibility the decisive design questions.
Why it matters
For evaluating an offering this is the core question: is a flowchart being sold here, or a system that plans for itself? Either can be right. A fixed workflow is checkable, repeatable, and cheap, and for processes with clear rules almost always the better choice. An agent pays off where the cases are too varied for a diagram. The price is that you can no longer test it thoroughly, only observe and limit it. Whoever builds agentically where a fixed workflow would have sufficed buys fuzziness with nothing in return.
Commonly misunderstood
“Agentic” has become a sales word. The useful test question: which action does the system carry out without asking, and how do you undo it?
Where this term does its work in the magazine.
2 articles use the term or take up its principle.
From Science
Benchmark Performance ≠ Bedside Performance
A good average can hide precisely the cases in which a system becomes dangerous.Read the articleFrom Science