Strip the mystique and an agent is a task that survives interruption.
A prompt answers and forgets. An agent keeps a plan, works through it over minutes or hours, and can be stopped, inspected, and resumed. That durability is the whole trick, and everything hard about agents follows from it.
The loop, plainly
The agent holds a goal and a list of steps. It takes the next step, observes what happened, updates the list, and decides what comes next. Sometimes it asks a question. Sometimes it stops. That is the entire architecture; the model inside decides how well each turn goes, but the loop is what makes it an agent rather than a query.
Property one: a plan you can read
Before work begins, the plan should exist in plain language. A user who can read step three and object is a user who will trust steps four through twelve. Plans that only exist inside the model are plans you cannot audit.
Property two: durable state
If the process restarts, the work must not. Each completed step has an artifact: a draft, a record, a delivered message. Resumption picks up from the last artifact, not from the beginning.
Property three: checkpoints on the expensive verbs
Agents that send, pay, or delete without pause are not autonomous, they are unattended. The working pattern gates the costly actions behind approval and runs everything else at full speed.
Judge any agent product against those three properties and the conversations get much shorter. Does it show the plan? Does the work survive a restart? Where does it stop and ask? If those answers are clear, the rest is implementation detail.