Foundations · Chapter 03
Agency
Agency is not automation. It starts when a system compares futures before it acts.
The capacity of a system to form counterfactuals (models of possible futures that don't exist yet), use internal models of itself and the world to predict where each action leads, and act purposefully toward the futures it prefers. A thermostat reacts to the present: that is control, not agency. Agency starts when a system compares possible futures before it acts. Glossary ↗
A system can carry out a great deal of work without choosing between possible futures. A fixed workflow can route a request, send a message and update a record. Its usefulness does not depend on calling it an agent. The important distinction is whether the next step is prescribed or chosen using a model of what could happen.
For a business, that model needs somewhere to live. The fact layer describes the business and its world; the hypothesis tree records the futures and bets it is testing; agents give it a way to act. These parts work together. Action without a reliable model is difficult to steer, while a model that never changes an action does little practical work.
Capability and permission
An agent’s ability to complete a task does not decide what it should be allowed to do. Autonomy is a deployment decision: which actions can happen without another approval, which require a person and which are forbidden. A capable system can be useful inside a narrow boundary.
To assess a system, ask how it chooses its next step, whether it predicts where an action will lead and what it may do without asking. Then identify who set that permission and who answers for the result. These questions separate reasoning capability from delegated authority.
Where to introduce autonomy
Start with a fixed workflow when the path is predictable. Give a system freedom to choose its route where the task contains uncertainty that a fixed path cannot handle well. The purpose is to solve the task with a level of independence that can be supervised, rather than to maximise the amount of autonomy on display.
A useful operating record names an owner and an autonomy level for every agent in production. It should also make clear which processes remain fixed workflows. Shared language matters here: if every automated step is called an agent, it becomes harder to see where the business has actually delegated a decision.