Agents and agent mode
An agent is the smartest way vragen.ai answers questions, and the direction the product is evolving in. An agent doesn't answer in one go, but takes multiple reasoning steps: it determines what it needs, searches your knowledge base in a targeted way and builds its answer from that. For harder questions this produces better and better-substantiated answers.
In the article From a search query to a reasoning agent we show visually how an agent reaches an answer.
A new agent starts with a conversation
Creating an agent is not a form full of settings but a short conversation in plain language: what is the agent for, for whom, which tone fits, and what must it absolutely not do? The clearer your answers, the sharper the next step becomes.
Based on those answers, the AI proposes concrete rules. "Never promise prices that are not in the knowledge base" becomes, for example, the rule "No unconfirmed commitments", with a written-out instruction you can read and adjust right away. You refine, the AI does the writing.
Rules: how you steer your agent
Rules are the core of self-reliance: they state in plain language what an agent may and may not do, and you manage them yourself. The Behavior tab shows four kinds:
-
System rules come from vragen.ai and are on by default; you cannot switch them off. They safeguard the baseline quality of every agent: correct source citation, sensible search behavior, staying on topic, and the right answer format. Time awareness lives here too: the time-context system rule teaches the agent to handle day and date, so "tomorrow" in an answer really is tomorrow.
-
Standard rules are ready-made rules from vragen.ai, for example for tone and answer length. They arrive through the wizard, or you link one yourself via New rule and the "Based on" field. You switch them on or off, and Detach and edit turns one into your own version. The rule remembers what it is based on, so Use example again always takes you back to the original.
-
Organization rules apply to all of your organization's agents at once, for example a fixed house-style rule. An Owner manages them in one central place; every agent receives them automatically and can deviate per rule.
-
Own rules belong to a single agent: a title, an instruction, done. The title is not just for you; it also helps the agent understand the rule.
Improving rules with AI
You do not have to write rules by hand. The Improve rules with AI button first asks what you want to change ("make the tone more formal", "add a rule about answer length") and then gets to work: the AI reads your existing rules, only touches what your request requires, and marks per rule what is new, changed or redundant. None of it is saved until you have reviewed it and click Save rule; close the screen without saving and nothing has changed.
It also works the other way around: Generate from rules writes a short description of your agent based on the rules that are already there, so the name and description keep matching the behavior.
What does an agent do extra?
Where the classic system runs one search and answers based on that, an agent can take steps of its own while reasoning:
-
split a broad question into subquestions and look them up separately;
-
search multiple times if the first results aren't good enough;
-
combine information from different sources into one answer.
An agent sticks closely to your instructions and stays close to your own sources. That makes the answers predictable and well substantiated.
An agent doesn't always have to think longer. You can make it work just as simply and quickly as a classic system, and for more complex questions it uses its extra reasoning steps. Which settings determine the speed is covered in How fast is an AI agent?
Configuring the agent in the app
In the Agents module you tune every agent yourself, without code. Open an agent from the list and you land in its settings, spread over four tabs: Behavior, Interactions, Model and Advanced.
The most important controls per tab:
-
Behavior: the agent's name, description and rules. Standard rules (such as a tone of voice) can be switched on or off, edited yourself or sharpened with AI; your own rules decide how much room the agent gets outside your knowledge base.
-
Interactions: the visual elements the agent can show during the conversation, which you design in the Widget studio, and the MCP connections it may use (see below). Visual elements come with the Essential plan and up, MCP connections with Professional; on a smaller plan you still see the tab, with a note on what it takes. If website tools are switched on for your environment, you set those up here too.
- Model: pick the language model and the reasoning level, from answering directly to extensive reasoning on hard questions. Every model shows its trade-off between intelligence and speed, and whether it is hosted in the EU; the latter can matter for your data agreements.
- Advanced: connect a search strategy, let the agent fetch relevant context at the very first message with Preflight search (which makes it more focused and faster), and restrict which sources it may use.
An environment can have several agents side by side, each with its own settings; you connect them per integration and appoint one as the default. With Try out you test a change immediately, before saving it, with Duplicate you start a variant from an existing agent, and with Add to playground you set an agent up for your team in the playground. Managing agents starts at the Editor role.
Switching without losing your work
Already done optimization rounds on your current system? They're not lost. When moving to an agent, your instructions and tweaks come along. A technique like splitting up questions keeps working too; an agent just does it in its own way, as one of its reasoning steps.
The safest route is trying it out before you go live. In the playground you select the agent at the top and compare its answers with those of the classic system. Once you're satisfied, we put the agent live for you.
Agents are the future of vragen.ai: by the end of 2026 all environments will use them and we'll be migrating existing systems. So switching is the route either way, and as described above your optimization work isn't lost in the process.
What an agent may use
Per agent you configure which part of your knowledge base it may consult. That's useful if you have multiple websites or languages in one environment:
-
Give each agent access to only its own content, for example a Dutch and an English agent side by side.
-
Or let all agents access everything, and use instructions to steer how they handle the different languages or brands.
Which approach fits depends on how your content is structured. We're happy to think along. See also Managing content in your knowledge base.
Connecting external data sources (MCP)
Besides your own knowledge base, an agent can also consult an external data source via an MCP server (Model Context Protocol). Think of your own vector database or an external API.
You add the MCP server in your environment and then switch it on or off per agent. That way you decide which agent may use the extra source. If the agent cites something from such an external source, it shows up as an external source next to the answer, with a link to the page itself; in your conversations those sources are listed apart from your own knowledge items, so you always know where an answer came from.
This is the reverse direction of Knowledge base as an MCP server: there you offer your knowledge base as a source to an external AI client such as Claude. Here your agent uses an external source.
This is how we offer LexAgent, for example: a source with Dutch laws and regulations, local regulations (such as a municipality's general local bylaw), official publications and court rulings. LexAgent can also be used on its own via lexagent.nl. Want to do something with this, or curious whether we can offer a particular source as MCP? Get in touch via service@vragen.ai.
Designing visual elements in the Widget studio
An answer does not always have to be text. An agent can also show a visual element: a card with opening hours, a product comparison, a step-by-step plan with buttons. You design those elements yourself in the Widget studio, on the Interactions tab.
You don't need to write code for it. In the studio you describe to an AI assistant what you want to see, and it builds the element for you on a canvas, with sample data so you see right away how it looks. Not quite right? Say what should change and the assistant adjusts it. Your work only goes live when you save the agent.
A good-looking element is only useful if the agent knows when to show it. That is what the Teach this agent how to use this widget button is for. You write in one sentence what you have in mind, for example "Show the opening hours when someone asks when we are open", and the AI writes or adapts the agent's rules. You see that proposal before it takes effect. If an element is not mentioned in any rule yet, the list labels it Not in the rules yet, so no design sits unused.
When the agent shows several elements in a row, you choose per element how that looks: one below the other, as a carousel with horizontal scroll, as a stack with arrows, or side by side, wrapping onto a new line. Made an element another agent could use as well? Copy it and paste it into that agent. The Widget studio is included from the Essential plan.
Actions on your own website (website tools)
An agent can also do something on your website, in the visitor's browser: find a product in your own catalog, or book an appointment. That works with WebMCP, an open standard in development that lets a web page state which actions an AI assistant may perform on it. Your website offers the actions, the agent uses them.
You stay in control. Per agent you list which websites are trusted; only pages on those websites can offer actions. Every action the agent comes across on your websites shows up on the Interactions tab, with a label such as Read-only or Consequential. Per action you switch it on or off, adjust the description for the AI, or choose Always ask the visitor first. If your website itself marks an action as consequential, the widget always asks the visitor for permission first. Without that approval, nothing happens.
Website tools are new, and we switch them on by request, together with you, because your developer first has to register the actions on your website. Want to get started? Email service@vragen.ai. If you run the widget behind your own proxy, also read The widget in a locked-down environment.
Which language model does an agent use?
An agent works with the language model you choose in the agent settings. How much choice you have there depends on your plan: on Starter your agent runs on one fixed model in the EU, from Essential you pick from a selection of smarter models, and from Professional from every available model. If you see a model you cannot select, it says which plan it needs. GPT-6 is among them, hosted in the EU: the fast version from Essential, the heavier one from Professional. More about this in Hosting and data storage.
Language models are succeeded quickly. When a model will be replaced soon, the model picker labels it Will be replaced soon, and an agent running on it shows a notice with a Pick a model button. That way you decide when to switch. If a model does disappear before you have chosen, your agent simply keeps answering on a model in the same region. Your visitors notice nothing, and your data stays where it was.
Want less choice instead, because your organization only allows Dutch or European models? Then we pin your environment to that selection as custom work: only fully sovereign models on our Dutch servers, or only models that run within the EU. No one in your environment can pick another model after that, not even by accident.
Want to know more?
Want to switch, or first see what an agent delivers for your situation? Email service@vragen.ai and we'll set up an agent for you to test. See also Testing and evaluating to assess the answers.