The Agent That Can Act

An AI system that can take steps across services and institutions raises different risks from one that only answers a question.

The paired portraits and credits appear on the series home. This is Martin’s independent commentary on Kohler’s published work. Alan Kohler has not reviewed or endorsed it.

Kohler’s documented work

What he has argued

Kohler has focused on the defining feature of agents: they can act with some degree of autonomy. His October argument describes the productivity possibilities and the dangers that follow when systems make decisions, access tools or affect infrastructure. He highlights cybersecurity and human dependence, and says there is a strong case to ban autonomous AI. That is his published position; this series reports it without presenting the proposal as a settled consensus.

Martin’s response

Where I agree, and where I go further

The word “agent” can make a technical capability sound harmless. The real questions are practical: what permissions does it have, whose interests does it serve, what can it change, and who is answerable if it causes harm? I think control should be discussed in terms of concrete powers and consequences, not only the sophistication of the model.

Open question

What I still want to understand

Should an agent be allowed to act only within a person’s explicit boundaries, or can any system be trusted with broader authority? What kinds of actions should require human approval, and what happens when the human is unavailable?

Research trail

Sources

Published 18 October 2026 · Commentary by Martin Davies-Roundhill · Contact End Up Here