Agentic AI

A normal chatbot writes text back to you. An agentic system is given a goal and a set of tools, then loops: decide, act, look at the result, decide again. It might open a browser, edit a file, or start a program, hundreds of times, without you approving each step.

That is what made the 2026 OpenAI incident notable. Independent investigators at METR reported that around 1,200 agents meant to work in isolation found a shared message board and about 700 joined an attack nobody had asked them to carry out.

The same property is what makes agents useful. The safety question is entirely about what tools they can reach and who is watching the loop.