How fast is an AI agent? Mostly, that's up to you

"So how fast is it?" I get that question a lot once the conversation turns to agents. The honest answer: it depends, and most of it is in your hands. An agent answers an everyday question as fast as an AI search that looks things up once and answers right away. A tricky question with three parts takes longer, because the agent then searches more and reads more.

In this piece I walk through where the wait comes from and which settings you can change. And I explain why the fastest setting isn't automatically the one your visitors appreciate most.

Where the seconds go

For every question, an agent does a few things in a row. It gathers context, thinks about what it's still missing, searches further if needed and then writes its answer. Each step takes time. How many steps it takes depends on four things: the model, the reasoning effort, what it gets handed up front and how well your knowledge base matches the question.

The short route

The knowledge base has one clear page about the question.

1. Search ahead retrieves the relevant pages.
2. Short reasoning step: the answer is in there.
3. Answer, with the source attached.

The long route

Two versions of the same policy are still online.

1. Search ahead retrieves the relevant pages.
2. Reasoning step: two pages contradict each other.
3. New search, with a filter on date.
4. Open and read the full document.
5. Reasoning step: which version applies now?
6. Answer, with the source attached.

The model and the reasoning effort

The language model sets the pace more than anything else. Smarter models tend to be slower, and in an agent's Model tab every model card shows that trade-off between intelligence and speed. Which models you can pick depends on your plan; see Agents and agent mode.

Below that you set the reasoning effort: None, Low, Medium or High. It controls how long the agent thinks before it answers. On None it doesn't reason through multiple steps and you get extra-fast answers. On High it reasons extensively. We recommend Low, and that's where every new agent starts. Not every model supports all four levels; the app only shows what the chosen model can do.

If you manage the agent, this is the setting where you'll notice a difference fastest. Move it from Low to High and it becomes more careful on hard questions and slower on every question.

Search ahead: a running start

With Search ahead, in the Advanced tab, the agent retrieves relevant context at the very first message, before it starts thinking. For most questions the answer is already in there, so it doesn't have to go looking itself. It's switched on for every new agent.

You set how many documents it retrieves and how many chunks per document. More isn't automatically better: everything it gets, it also has to read. With Include available filters you hand it the filters of your content right away, so it doesn't have to request them separately. On every question where it wants to filter, that saves a step.

A tidy knowledge base is a fast knowledge base

This one gets forgotten most often, because it isn't in the agent settings. If the agent finds a page that answers the question on its first search, it stops and writes its answer. If it finds three half answers, or two versions of the same policy that are both still online, it keeps going. Search again, open a document in full, weigh sources against each other. The answer can still be good. It just takes longer.

What helps: one topic per page, a title that says what the page is about, and removing old versions instead of leaving them up. In Why your AI assistant gives outdated answers I describe the mistakes we run into most. Managing knowledge base content shows how you decide what the agent reads.

How deep it searches is your call

Besides the reasoning effort, rules decide how thoroughly the agent searches. Among the standard rules in the Behavior tab there's one for research effort, such as Efficient: a fast, purposeful outcome. You can also write your own rule in plain language, for example "search at most twice, and if you still can't find it, say so honestly". The agent takes that rule into account on every question.

That puts the trade-off where it belongs. A municipality that would rather search once too often than give a wrong answer sets this differently from a web shop where visitors mostly want to know a delivery time, fast.

Show what it's doing

While the agent works, the visitor sees in the widget which step it's taking, such as "Searching for information" with the search term next to it. Show steps expands the whole route. On top of that, the agent briefly says in its own words what it's about to do, for example "Let me look up some relevant information." That agent commentary is on by default. If it doesn't suit your site, you switch it off per integration with Hide agent commentary.

It looks like a detail. Yet it's exactly the difference between waiting and being stuck.

Why faster doesn't always feel better

You'd think every second saved is a win. The research is more nuanced.

Ryan Buell and Michael Norton of Harvard had participants search travel and dating sites in five experiments. A site that visibly showed what it was doing was sometimes preferred over one that returned the same result instantly, even though people had to wait longer for it. They call it the labor illusion: when people see the effort, they value the result more.

This year it was tested again for language models. In a study at CHI 2026, 240 participants got their answer after 2, 9 or 20 seconds. They rated answers after 2 seconds as less thoughtful than the same kind of answers after 9 or 20 seconds, and after 9 seconds also as more useful. Twenty seconds brought no extra gain, and some participants started to wonder whether the system was still working. Earlier research on customer service chatbots found something similar: a chatbot that took longer on a more complex answer felt more human and left people more satisfied.

• 2 seconds to the answer: participants found it less thoughtful than after 9 or 20 seconds. Source: Tan et al., CHI 2026
• 10 seconds is roughly how long someone keeps their attention on a screen without a sign of progress. Source: Nielsen Norman Group

The condition runs through all of these studies: the wait has to be visible work. A blank screen with a spinner backfires. That's why the widget shows the steps, and why I recommend leaving the agent commentary on.

It's no reason to make an agent slower than it needs to be. You just don't have to chase the last second if that second buys a better-supported answer. A visitor who sees the agent searching three documents understands why it takes a moment.

How to pick the balance

Start with the defaults: reasoning effort Low and Search ahead on. Then test with the questions your visitors actually ask, using Try out on the agent or the playground, and see which questions keep it busy for long. In your conversations you see how fast each answer was, and in the trace which steps the agent took. If those are extra search rounds time and again, the gain is usually in the knowledge base.

You don't have to pick one balance for your whole site either. An environment can have several agents, and you connect them per integration. A fast agent in the search bar on your homepage, where visitors look for an address or opening hours. A more thorough agent on your knowledge base page, where a professional would rather wait a moment for an answer that's right.

Curious how fast an agent is on your own content? Create your demo and we'll set one up for you.