Beyond the answer
Agents connect language and inference to action: using tools, working within environments, delegating tasks, and adjusting to feedback.
The environment matters as much as the model. A useful agent needs a way to act, a way to observe the effects, and boundaries that make experimentation possible.
The shape of an environment
Safe sandboxes, tool use, reinforcement learning, and autonomous workflows sit close together here. Each changes the relationship between a system and the world it can affect.
Local runtimes add another question: what happens when this capability belongs to the person using it? Sato is one place to work through that question.
A working entry. The connections may change.