Donné Stevenson found that dealers rarely engaged with an open chat box because they did not know what to ask or what the agent could do.
2
Preset buttons, page-specific actions, streaming, and interactive answers helped users move from a first click to follow-up questions.
3
Purpose-built aggregation tools, summary statistics, CSV data, and compact clickable tokens made the agent's data access more manageable and less expensive in context.
Summary
Donné Stevenson describes building an agent for car dealers on Autooto, a Polish secondhand-car marketplace. The aim was to help dealers understand their listings and take action, rather than answer general FAQs. The first version reached all users within two weeks, but only 10% engaged in a measurable way and repeat use was negligible. Dealers did not know what to ask, while preset question buttons received more interaction than open questions. The next version added a dynamic toolbar, page-specific actions, streaming responses, and interactive links. Stevenson explains the trade-offs behind purpose-built aggregation tools, including reliable but limited tools instead of unrestricted SQL generation. The team also moved from JSON to CSV with summary statistics to reduce token use. The talk argues that agent intelligence needs a usable interface around it. Buttons can provide the first step, after which users are more likely to ask follow-up questions.
Stevenson says users have to understand what the agent can do, how to ask for it, and how to use the answer. For dealers who are already busy and overwhelmed, the blinking cursor creates another task. The first product therefore focused on helping dealers manage listings and use the platform's data, rather than presenting another FAQ assistant. The agent was meant to read business information, suggest what to do next, and reduce the work needed to organize listings and act on insights.
The first release was a basic React agent with a few tools for data retrieval and analysis. It allowed users to ask about any feature, even when it could not always answer, because the team wanted feedback quickly. Within two weeks, the agent had been shown to 100% of users. Only 10% of dealers engaged in a measurable way, and repeat usage was negligible. Users often asked what the agent could do, became frustrated by its limits, and clicked suggested questions more often than they typed their own.
The first experiment showed Stevenson that users wanted insights but did not know how to start. Suggested questions worked better when users could click them, although engagement was still low. The next hypothesis was to ease users into the agent without formal onboarding, training, or long tooltips. The team added buttons and a dynamic user interface so the first interaction could begin with a concrete action instead of an empty chat field.
The second version placed chat alongside other actions in a navbar. Some buttons, such as upload, sell, and extend, were ordinary front-end shortcuts. A recent-changes button opened the assistant with a preset question about inventory movement, including sold listings and ads nearing expiration. The navbar changed with the page. On an announcements page it could offer filters for ads that might expire, while on an inquiries page it could show messages needing a reply. Stevenson says these buttons received substantial usage because they anticipated actions users might want to take.
Purpose-built aggregation tools balance safety and flexibility
Stevenson describes a spectrum in tool design. Highly narrow tools are reliable and safe but cannot handle requests outside their exact purpose. A tool that lets the agent write arbitrary SQL is flexible but makes tool choice and context much more complex. The team settled in the middle with a fixed set of purpose-built aggregation tools. Each tool targets a specific concept, aggregates data for that use case, explains terms in plain text, and returns a smaller raw-data sample for follow-up needs.
Page context can be approximated without building a web agent
Dealers expected the assistant to know what they were viewing, but the team was not building an agent that could see the whole user interface. Full context awareness would have been a large project. Instead, the navbar changed according to the active tab and supplied preset functions relevant to that page. Stevenson says this gave users the experience of page awareness without adding the full complexity of making the agent understand the entire interface.
Streaming and data representation reduce the cost of waiting
The agent's P99 response time was almost 20 seconds. Streaming did not make the model smarter, but it made the response begin sooner and gave users more reason to continue. The team also changed how retrieved data was represented. JSON was easier to interpret because each record contained its own labels, but it used almost twice as many tokens as CSV. They moved to CSV and added summary statistics at the top, accepting some loss of readability to reduce context cost.
Stevenson says plain text is difficult to consume when there is a lot of it, and it does not give users an immediate action. The interface therefore replaces parts of the response with dynamic elements. A reference to an advert becomes clickable, taking the dealer to the full advert. The team first returned full links from the agent, but URL-encoded links consumed many tokens and could be malformed. They replaced them with compact tokens containing an advert ID and name, which the front end converts into a link.
In the results Stevenson discusses, users rarely opened a session with an open question. When they clicked a button first, they were more likely to ask another question later. The team treats button clicks mainly as engagement events, while follow-up questions provide a separate signal about the agent's responses. Stevenson expects the interface to become more dynamic and to anticipate more of what users want next. Personalization is another unresolved issue because the users are businesses and the team is still working out what personalization should mean for dealers.
"The biggest learning that we've taken from it is the idea that as we're moving more and more into an agentic world, we need to move past just a chat interface and start looking at more of the instant action and more of a dynamic experience with these smart agents."00:04
Who should watch
You are building an agent for users who are busy, unfamiliar with AI tools, or unsure what to ask.
Your agent returns business data and you need to decide how much aggregation, raw data, and tool flexibility to expose.
You are measuring whether interface actions lead to deeper agent use and want practical examples of buttons, streaming, and interactive responses.
