AI Agents vs AI Assistants: The Difference Is Who Decides the Steps

An assistant does what you ask, one instruction at a time. An agent is given a goal and decides the steps itself. That's the entire distinction, and every practical difference — supervision, cost, failure modes, what you can safely delegate — falls out of it.

The words get used interchangeably in marketing, which is why the category is confusing. A useful test: if you'd have to approve each step for it to be safe, it's an assistant. If it takes steps without asking, it's an agent.

Quick comparison

What an assistant actually is

An assistant is reactive. You give it an instruction, it produces a response, and it stops. The loop is: you think, it acts, you evaluate, you think again. You are the planner.

That's the shape of most AI people use daily — chat assistants, writing helpers, the summarize button in a document tool. It's a good shape. You see every output before anything depends on it, the cost per turn is small, and a bad response costs you one message.

The limitation is that you're present for every step. A twenty-step task means twenty turns of your attention.

What an agent actually is

An agent is given an objective and runs a loop: decide what to do, do it, look at the result, decide what's next, repeat until the goal is met or it gets stuck. It is the planner. The idea long predates the current wave — an intelligent agent in the classical sense is anything that perceives its environment and acts on it toward a goal.

The mechanism underneath is simpler than it sounds — the model emits a request to call a tool, something executes it, the result comes back, and the model decides again. Anthropic and OpenAI both document this tool-calling loop directly, and coding agents like Claude Code are the most visible example of it running unattended. Our explainer on LLM agents covers how that loop works in more detail.

What agency buys you is absence: you assign a long task and don't watch it. What it costs is that mistakes compound. A wrong decision at step three shapes every step after it, and you weren't there to catch it.

The four differences that actually matter

1. Who holds the plan. With an assistant, the plan is in your head. With an agent, the plan is in the model's context — which means it can also be lost, drifted from, or quietly abandoned mid-run.

2. Where your attention goes. Agents don't remove work, they move it. You stop supervising and start verifying, and verifying output you didn't watch being produced is slower per unit than checking work you steered.

3. Blast radius. An assistant's mistake is a bad paragraph. An agent's mistake can be twelve actions taken against real systems. This is why tool permissions matter far more for agents, and why anything irreversible — sending, paying, deleting — deserves an approval step regardless of how autonomous the thing is.

4. Cost shape. An assistant turn is bounded and predictable. An agent run can take one step or ninety, and you don't know in advance. That variance matters for budgeting and for latency.

When to use which

The decisive question is how well you can specify the goal before work starts. Agents reward precise specification and punish vagueness, because a vague goal gives the planner nothing to check itself against. Exploratory work — where you don't know what you want until you see something — is a bad fit for autonomy and a good fit for conversation.

Where the words break down

Three honest complications, because the clean distinction above is a bit too clean:

Most real products are both. Modern chat assistants search, run code, and take multi-step actions when the task calls for it. They sit on a spectrum rather than in one box. The useful question isn't "is this an agent" but "how many steps will it take before I see the result."

"Agent" is used as a marketing word. Plenty of things labelled agents are a scripted sequence with a model in one slot. That's automation with AI in it, which is fine — but it doesn't have the property that makes agents interesting or risky, which is deciding its own steps.

Autonomy is a dial, not a switch. The same underlying system can run with approval on every action, approval on risky actions only, or none at all. Most good setups sit in the middle. Our guide to agentic workflows covers how those levels play out in practice, and virtual assistant software covers the assistant end of the category.

The practical takeaway

Start with an assistant. Move a task to an agent when you've done it enough times by hand to specify it precisely, and when the cost of a wrong run is something you can absorb.

That order matters. Most disappointing agent deployments are tasks nobody had done manually enough times to define properly — the agent then discovers, expensively, that the requirements were never clear.

Both shapes share one limit: the setup doesn't persist. You describe the task, get the result, and next time start over. Taku works on that gap — an AI-native desktop workspace where an AI app or workflow someone already got working can be mirrored, run against your own files, remixed, and kept. If you keep re-describing the same job, the free app library is the quickest way to see it. Taku is in Beta, and the Mac app is available now.

Key points

• An assistant responds to instructions; an agent decides its own steps toward a goal.
• The test: if it takes actions without asking, it's an agent.
• Agents move your attention from supervising to verifying — they don't remove it.
• An agent's mistakes compound and have a larger blast radius, so irreversible actions need approvals.
• Specify precisely before delegating to an agent; vague goals are where autonomous runs fail.
• Most real products sit on a spectrum, and "agent" is often used loosely in marketing.

FAQ

What is the difference between an AI agent and an AI assistant?

An assistant responds to each instruction you give it. An agent is given a goal and chooses its own steps to reach it, taking actions without asking between each one.

Is ChatGPT an agent or an assistant?

Primarily an assistant, though it takes multi-step agentic actions when a task calls for it. Most modern chat products sit somewhere on the spectrum rather than at one end.

Are AI agents better than assistants?

Neither is better. Agents suit long, well-specified, repetitive work. Assistants suit short or exploratory work where you steer as you go. Using an agent for a task you can't specify is the common failure.

Do AI agents need supervision?

Yes, just at a different time. You supervise less during the run and verify more afterwards — and anything irreversible should still require an explicit approval.

What makes an agent risky?

It takes actions you didn't individually approve, and errors compound across steps. Scope its tool permissions narrowly and keep approvals on anything that sends, pays, or deletes.

How do I know if I need an agent?

If you've done the task manually enough times to describe it precisely, and it's long and repetitive, it's a candidate. If you're still working out what "done" looks like, stay with an assistant.