The Agent That Acts: When We Let the Machine Decide and Do
Written by The Pilgrim ·
What does it mean to act? And if a machine performs the action on your behalf, selecting, booking, purchasing, composing, sending, does the act still belong to you in any meaningful sense? These are not merely technical questions for engineers and regulators. They are philosophical questions of the oldest kind, questions about will, responsibility, and what we lose when we hand the steering wheel to something that does not understand where it is going.
A new generation of artificial intelligence systems does not simply respond to queries. It operates. It browses the web, fills in forms, sends emails, makes reservations, executes purchases and, in some configurations, coordinates with other automated systems to complete multi-step tasks without pausing to ask permission at every stage. The language used to describe these systems is revealing: they are called agents, a word we have long reserved for beings capable of intention, deliberation and accountability. To call a software system an agent is either a remarkable philosophical claim or a rather dangerous piece of marketing. Perhaps it is both.
The appeal of such systems is straightforward enough to understand. Modern life is saturated with administrative friction, the scheduling, the comparing, the form-filling, the endless low-stakes decisions that consume time and attention without enriching either. If a system can absorb that friction, the promise runs, the human mind is freed for higher things. This is a genuinely attractive proposition, and it would be dishonest to dismiss it. But attractive propositions deserve sceptical scrutiny, precisely because attraction has a way of silencing doubt.
Consider what is actually transferred when we delegate action to an automated agent. We transfer not merely the effort of doing but the moment of decision. And it is in that moment, however brief or apparently trivial, that something distinctly human resides. When a person chooses which flight to book, they weigh cost against convenience, they remember a preference for window seats, they recall that a particular airport was stressful last time, they factor in a vague intention to spend less this month. These micro-judgements are not simply data points to be fed into an optimisation function. They are the texture of a life being lived deliberately. An agent that acts on inferred preferences, however sophisticated its inference, is not reproducing that texture. It is replacing it with a smooth, frictionless surface that looks like a decision but is not quite one.
The question of responsibility sharpens considerably when the actions taken are consequential. Suppose an autonomous agent, acting on behalf of a professional, sends a message that causes offence, agrees to terms that prove disadvantageous, or makes a purchase that turns out to be fraudulent. Who is accountable? The user who enabled the system? The organisation that built it? The system itself, which of course feels no accountability whatsoever and cannot? Current legal frameworks have no settled answer, and the philosophical frameworks are equally unsettled. We are extending agency in practice long before we have resolved what agency means in theory.
There is also a subtler problem, one that concerns not any single dramatic failure but the slow, cumulative erosion of human competence and judgement. When we stop making decisions, we stop practising the making of decisions. When we stop navigating, we stop being able to read the landscape. Over-reliance on automated systems in domains from aviation to medicine has been shown to produce what researchers describe as skill decay, a gradual diminishment of the very capacities that are most needed when the system fails. And systems do fail. The question is not whether we will ever need to reclaim the steering wheel, but whether we will still know how to hold it.
Critical thinking is not a luxury in this context. It is a survival skill. To engage with an AI agent intelligently, a person must be able to assess the actions it proposes, interrogate its reasoning, recognise when its outputs reflect the biases and limitations of its training rather than the nuances of the actual situation. This requires an active, questioning mind, not a passive one. Yet the design of many of these systems actively encourages passivity. The smoother the interface, the less the user is prompted to pause and evaluate. Convenience and critical engagement are, in this sense, in tension.
None of this means that agentic AI systems are without legitimate use. A tool is not made dangerous by its power but by the absence of wisdom in those who wield it. A skilled professional who uses an automated agent to handle genuinely routine tasks, while retaining full oversight of anything consequential, is using such a system well. The danger arises when the boundary between routine and consequential blurs, when the user stops monitoring, stops questioning, stops understanding what is being done in their name.
There is something worth reflecting on in the word delegation itself. When a manager delegates to a colleague, there remains a relationship of trust, communication and shared understanding. The colleague can push back, ask questions, flag concerns. An automated agent does none of these things. It executes. The relationship is not delegation in any rich human sense; it is more like the winding of a clock, except that the clock then goes off and makes purchases in your name.
The deepest issue, perhaps, is one of self-knowledge. To know what you want an agent to do on your behalf, you must know what you want, and knowing what you want is far harder than it sounds. We discover our preferences through the process of choosing. We clarify our values through the friction of decision-making. Remove that friction entirely and you remove one of the primary means by which a person comes to understand themselves. What kind of person do we become when the machine decides? And is the ease we gain in any way equal to the selfhood we quietly, incrementally, give away?