# The difference between an AI assistant and an autonomous agent | Mattia Ciuni

> AI assistant vs autonomous agent is not a question about capability. It is a question about who is present when the irreversible thing happens, and you can answer it in one sentence by pulling the network.

- URL: https://mattiaciuni.com/thoughts/the-difference-between-an-ai-assistant-and-an-autonomous-agent
- Type: Blog post
- Published: 2026-11-19

## Full article

# The difference between an AI assistant and an autonomous agent

Most of the argument about **AI assistant vs autonomous agent** is conducted with adjectives. I want to replace it with one sentence: an assistant stops when the human is absent, an agent does not. Everything else follows from that.

## The sentence, applied to things that shipped

Google's buying surface, launched in January 2026, is the clearest example. The agent builds the cart, negotiates, fills in what it can, and then the documentation says the user interacts only with the interface to submit the order, with the agent not involved in that part. Single item, human at the end. That is an assistant, and it is a very good one, and it is shipping at enormous scale.

OpenAI's Instant Checkout, announced in September 2025, is described the same way: users explicitly confirm each step, payment tokens are authorised only for specific amounts and specific merchants, and the merchant still accepts or declines. Also an assistant. Notably, the announcement said single item purchases only, with multi item carts to follow, which is a detail that tells you the hard part was the last step, not the first one.

Google's Agent Payments Protocol, in the version donated to the FIDO Alliance in April 2026, added something the buying surface does not yet use. It calls it Human Not Present payments: the agent executes autonomously on the basis of pre authorised user instructions. That is the definition of the other thing. It is also, notably, a protocol feature rather than a checkout feature, from the same company, in the same year.

I find that pairing more interesting than any forecast. The protocol and the buying surface are the same organisation. One of them puts the human back in at the moment of payment to ensure determinism. The other specifies fully delegated payment. The industry is shipping the capability faster than it is willing to use it, and the reason is written in their own documentation.

## The test: pull the network

If the user is asleep, an assistant does nothing and the morning contains a notification. An agent does the permitted thing and the morning contains a record. You do not need a definition from anyone. Break the connection at the moment of the purchase and see what the system does with the pending request.

I have been on both sides of this. My own prototype currently stops and asks: an order came back as a request to confirm at 163 euro, and it left at 137 after I intervened. That is an assistant, and I label it as one. The version I am building is the one where the request does not stop, and the honest thing about writing this piece is that the difference between the two is not intelligence. It is a document I write before the run.

The word is doing marketing work, and I would like to name how. Autonomous is used as a synonym for impressive, the way a product that asks permission on every step can still call itself autonomous because it plans the steps itself. Once you separate the two claims, most of the category resolves into something much more ordinary: an agent with a good planner and a strict rule about when to ask. That is a good product. It is not the thing the word is selling.

## Why the distinction is economic, not semantic

- **Assistants scale with human attention.** Every delegated action consumes a minute of someone's time, so the ceiling on autonomy is the user's calendar.
- **Agents scale with the quality of the boundary.** If the boundary is written, signed and narrow, the number of actions a user can delegate is limited by how much they trust the boundary, not by how often they can click.
- **Assistants need a good model.** Agents need a good record. The scarce resource changes, and it is not the one the industry is optimising.

This is why the same product can be an assistant on Monday and an agent on Tuesday. Nothing about the intelligence changed. What changed is that a person sat down and decided what the agent may do without them. The transition is not technical. It is a document, and it happens at a desk.

You can also tell from the outside, without reading the code, by looking for three things a serious implementation has and a demo does not. A **policy** the user can read before they grant anything, in language they wrote or approved. A **revocation** control that is described in terms of what happens to pending requests, not only to future ones. And a **receipt that explains the reason**, not just the amount. A product with none of the three is asking for trust it has not built the record to earn, whatever its marketing says about being autonomous.

None of this is a compliment to the assistants. A confirmation step is a legitimate product and for a lot of purchases it is the correct one. My argument is only about where the line gets drawn and who draws it, because the current market has an odd habit of describing the assistant as the agent and charging for the difference in vocabulary.

## The uncomfortable middle

Most systems marketed as autonomous are assistants with better marketing, and the evidence is always the same: look at what happens at the boundary. If every action can be turned into a confirmation, it is an assistant. If some actions can be turned off and others cannot, then there is a boundary in the product and the product is worth having, whatever it calls itself.

So the useful question for anyone evaluating one of these is not whether the agent can act. It is: which actions did you decide are safe without me, and who signed that decision. The answer, when it exists, is the real product. The rest is a demo.

I am building the agent side of that line at [Noesia](https://withnoesia.com). The document on the desk is [what should an AI agent be allowed to do](/thoughts/what-should-an-ai-agent-be-allowed-to-do/), why the payment step is where assistants stop is in [why AI agents keep stopping at checkout](/thoughts/why-ai-agents-keep-stopping-at-checkout/), and the record that proves it acted rather than asked is in [what an agent authorization record should contain](/thoughts/what-an-agent-authorization-record-should-contain/). The short version of the trust argument is [the agentic economy is a trust problem](/notes/the-agentic-economy-is-a-trust-problem/).

Building [Know Computer](https://knowcomputer.com), a personal context layer for the AI era. Reply via [m@knowcomputer.com](mailto:m@knowcomputer.com).

- [Thoughts index](../thoughts.md)
- [Home](../index.md)
