liveBook · Manning

Agency

As explained in Machines that Think

Agency is the capacity to initiate actions and pursue goals as an agent.

Agency is the capacity to initiate action and regulate behavior in pursuit of goals one treats as one’s own. It includes forming goals, revising plans, and responding to consequences as circumstances change. Intelligence and agency are related but distinct: a system can reason or solve problems without generating its own goals. section 5.4.2 section 9.5.4

Human goals are not created from nothing. Needs, learned preferences, habits, culture, and environmental pressures shape what people pursue. Agency is therefore not pure self-origination; it is a layered capacity to pursue and redirect goals within constraints. section 9.5.4 section 5.4.2

Artificial neural networks operate within arrangements that supply their objectives, inputs, and stopping conditions. A model produces outputs when prompted and optimizes during training, but remains inert otherwise. Coherent, goal-directed behavior within a task is not by itself evidence that the model generated the goal or can decide what matters. section 5.4.2

Agentic systems can be built around language models using tools, memory, and external objectives. From a skeptical perspective, their agency is borrowed: designers, users, prompts, and surrounding software supply the goals, while success and failure are external evaluations rather than stakes the model experiences. This limitation changes the kind of intelligence involved without settling whether intelligence is present. section 9.4.3 section 9.5.4

Figure 9.14 Agency and intelligence. Humans usually combine agency and intelligence, making the two easy to confuse even though they are different capacities. Agency concerns pursuing goals as one’s own; intelligence concerns understanding, reasoning, planning, and solving problems.

The figure contrasts pursuing goals as one’s own with understanding, reasoning, planning, and problem solving. section 9.5.4

Conversational continuity

A system may sound warm, recall prior exchanges, and maintain a conversational persona. Those behaviors can make an interaction feel continuous, but they do not establish awareness, commitment, agency, or care in the relational sense. Users may nevertheless attach significant meaning to the exchange. section 8.5.4

The distinction matters for judgment and responsibility. Mistaking task-directed behavior for self-directed agency can lead to excessive expectations of autonomy. A system without self-originating goals cannot be morally autonomous or bear responsibility for its actions, even if it displays cognitive performance. section 5.4.2 section 9.5.4

How Machines that Think covers it

How deep it goes
Defines agency and contrasts human goals with externally supplied machine objectives, while considering a skeptical account of borrowed agency.
Examples
Uses prompted model outputs, tool-equipped agentic systems, and emotionally meaningful chatbot interactions to illustrate the distinction.
What you can do after
Evaluate whether apparent goal-directed behavior supports claims about autonomy and responsibility.
How it connects
Connects agency to intelligence, world models, selfhood, and moral responsibility.

Also explained in

Compare how 6 books explain agency →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Machines that Think. The section links point to the parts of the book it draws on.