Ask a person at a counter whether a product is in stock and they will say "one moment" and check. Nobody minds, because they can see the person checking. Ask an AI agent the same question and, until this week, the visitor saw a fixed "Thinking…" for the whole wait. Same work, much worse experience.
This post explains, without jargon, why an agent that looks things up takes a few seconds, what the chat now shows while it works, and the handful of choices that decide how long your own agent takes.
An answer is a series of small steps
An agent that only talks from what it was told in its instructions answers fast. The agents worth building do more: they check your knowledge, call your booking system, read a product feed or search the web. Each of those is a step, and each step is a short round trip.
- 01Read the questionand decide what to check
- 02Look it upknowledge, a tool or the web
- 03Read the resultand decide if it is enough
- 04Look again?only if something is missing
- 05Write the answerstreamed as it is written
The agent reads the question and decides what it needs. It calls a tool and waits for the result. It reads the result, which takes a moment of its own, and decides whether that is enough or whether it needs another lookup. Only then does it write. A question that needs one lookup is quick; a question that needs a search, then the details of two results, then a price check, is four rounds.
The platform's own preparation is not where the time goes. Checking access, guardrails and limits, finding the right knowledge and loading the tools all happen in parallel and take a fraction of a second. The wait is the lookups and the reading between them.
What visitors see now
On the chat page and in the widget, "Thinking…" is now a short line that follows the agent's work: "Looking that up…", "Checking the documents…", "Browsing the website…", "Searching the web…", "Preparing the results…" and, before it writes, "Reviewing what I found…". If one step runs past 15 seconds, the line changes to "Still on it — this one takes a little longer…" so nobody wonders whether the chat has stalled.
- The lines come from a fixed list, translated into every chat language. A visitor never sees the internal name of one of your tools or what was sent to it, which may include personal details.
- When the agent runs several lookups at the same time, the visitor sees one line, not a flicker of them.
- There is nothing to configure. Every agent shows the steps.
Showing the work does not make it shorter, but it changes how the wait feels. A visitor who sees "Checking the documents…" knows the agent is doing what they asked.
Knowledge or a live tool?
The biggest choice you make is where each fact lives. Knowledge is your documents and pages, already prepared for search: reading it is quick and does not depend on any other system. A live tool asks one of your systems at the moment of the question: it is always current, and it is as fast as that system is.
If a fact changes once a month, it belongs in knowledge. If it changes by the minute, it belongs in a tool. An agent that calls a live system to find your returns policy is paying for freshness it does not need.
Five ways to make your agent faster
Keep the toolset small
Every tool is a choice the agent has to make. Five tools with clear, distinct jobs get picked right the first time; fifteen overlapping ones lead to a wrong call and a second round. Detach the tools an agent does not need.
Describe each tool in one plain sentence
Say what the tool returns and when to use it: "Returns live stock for one product by SKU. Use it only when the visitor asks whether something is available." A clear description saves a guess, and a guess costs a step.
Make the system behind the tool quick
A lookup can only be as fast as the system it calls. If an endpoint takes many seconds, that is the visitor's wait. Cache what you can on your side and return only the fields the agent needs: a short answer is also quicker to read than a page of data.
Tell the agent when not to look
Instructions can save whole rounds. "Greetings and thanks need no lookup." "If the question is vague, ask one short question before searching." A broad search on a vague question often leads to a second, narrower one.
Watch it in Preview
Ask your agent the questions your customers ask and watch the progress line. Several "Looking that up…" steps for a simple question usually point to a missing fact in knowledge or a tool description that is not clear enough.
Fast is not the only goal
It is tempting to strip an agent down until it answers instantly. Resist it. An answer that took six seconds and checked the live price is worth more than one that took one second and guessed. The aim is to remove the steps that add nothing, not the ones that make the answer right. When you are not sure whether an answer is right, your visitors will tell you: the 👍/👎 buttons under each reply now reach your dashboard, as described in chatbot answer feedback.
Frequently asked questions
Why does my AI agent take several seconds to answer?
Because it is checking things before it answers. Each lookup in your knowledge, your systems or the web is a round trip, and the agent reads each result before deciding what to do next. A question that needs several lookups takes several rounds.
What do visitors see while the agent works?
On the chat page and in the widget, a short progress line such as "Looking that up…" or "Reviewing what I found…", and after 15 seconds on one step, "Still on it — this one takes a little longer…". The lines are translated into every chat language and never show your tool names.
Do I need to switch the progress steps on?
No. Every agent shows them on the chat page and in the widget, with nothing to configure.
Should I put information in knowledge or behind a tool?
Facts that change rarely, such as policies, hours and descriptions, belong in knowledge: quick and steady. Facts that change by the minute, such as prices, stock and bookings, belong in a tool that asks your system live.
What is the quickest way to speed up a slow agent?
Open Preview, ask a typical question and count the steps. Remove tools the agent does not need, rewrite unclear tool descriptions, and make sure the system behind any slow tool answers quickly and returns only what the agent needs.
