Agent vs Assistant
The two words get used interchangeably in product copy, and the difference between them is most of this subject. An assistant responds; an agent proceeds. Almost everything else follows from that one shift.
Where the line falls
An assistant is turn-based. You ask, it answers, and control comes straight back to you — it might search the web or run code to produce that answer, but the unit of work is a single exchange and you decide what happens next. An agent is handed a goal rather than a question and keeps control across many steps: it plans, acts, observes the result, and decides for itself whether to continue, change approach, or stop. The practical tells:
- Who chooses the next step — the human in an assistant, the system in an agent.
- Horizon — one reply, versus a loop that may run for minutes or hours.
- Stopping — an assistant is finished when it has answered; an agent needs an explicit termination condition, step limit and budget, or it will not stop at all.
Why the distinction is practical, not pedantic
It is a spectrum, and the useful consequence is operational. The moment a system takes steps you did not individually approve, you inherit a different class of problem: runaway loops, cost overruns, irreversible actions, and a run nobody watched. That is why agentic systems need tracing, approval gates, permission scoping and sandboxing while a chat assistant needs none of them. Asked properly, "is this an agent?" is really the question "does this need those controls?" — and that one has an answer.
Agent vs assistant contrasts two modes of LLM-driven system. An assistant responds turn by turn, with the user deciding each next step; an agent is given a goal and chooses its own steps, calling tools in a loop until the task is done. The distinction is who holds control flow, and it determines how much supervision, permission scoping, and error handling the system requires.
An assistant is like a colleague you message. You ask something, they reply, and then it is your turn again. You stay in the driving seat the whole time.
An agent is like handing someone a job and walking away. "Book the whole trip." They work out what to do first, make the bookings, notice a flight is sold out, pick another, and come back when it is done.
The real difference is who chooses the next step. That is also why agents need extra safeguards — a spending limit, a rule for when to stop, permission before anything that cannot be undone. Nobody is checking each step the way they would in a conversation.