Goal versus command
A command is 'write a follow-up email.' A goal is 'make sure this customer gets a response today.' The second requires deciding whether to text or email, checking whether someone already replied, and recording the outcome.
Agents live in the second category, which is why they need access to real information and real actions.
The loop an agent runs
- Interpret the goal and the available context.
- Choose a tool and call it with validated input.
- Read the result, including failures.
- Decide whether the goal is met or another step is needed.
- Stop, and report exactly what was done.
What makes an agent trustworthy
Constrained tools, enforced permissions, confirmation before irreversible actions, and a complete record afterwards. Autonomy without those four is not sophistication — it is an unlogged employee with no manager.