How the Agent Works

This article explains the mental model behind TrustYou Agent. Understanding how the Agent thinks, what it can and can't do, and how knowledge, features, and conversations fit together will help you configure it effectively.

The Core Principle

The Agent only knows what you give it. It has no access to the internet, can't browse websites on its own, and won't make things up. Every answer it gives comes from one of two sources:

Knowledge you've added to its Places (hotel info, policies, FAQs, documents, reviews)
Guest context from connected external sources like the TrustYou CDP (booking details, stay history, preferences)

If the information isn't there, the Agent is designed to acknowledge that it doesn't know or hand over to a human rather than improvise an answer.

Knowledge vs Features

This is the most important distinction to understand.

Knowledge: what the Agent knows

Knowledge is the content you add to Places as Resources. Think of it as the Agent's reference library. It includes things like:

Hotel descriptions and amenity lists
Policies (cancellation, check-in/check-out, parking, pets)
FAQs and local tips
Website content you've pointed the Agent to
Guest reviews and feedback

When a guest asks a question, the Agent searches through this knowledge to find relevant information. It uses semantic search, meaning it understands the meaning of the question, not just keywords. If a guest asks "Can I bring my dog?" the Agent will find your pet policy even if it never uses the word "dog."

The Agent searches across all the Places you've assigned to it, but only those Places. That's how you control what each Agent knows and talks about.

Features: what the Agent can do

Features are capabilities you turn on or off in the backoffice. They don't add knowledge. They add abilities. For example:

Human Handover lets the Agent route a conversation to your team when it can't help
Quick Actions adds shortcut buttons to the chat interface
Guest Context lets the Agent access a guest's personal data from connected sources

You decide which Features to enable for each Agent. The Agent can only use the ones you've turned on. It can't decide on its own to enable a Feature.

A simple way to think about it: Knowledge is what the Agent reads. Features are what the Agent does.

There's a third piece working alongside those two: the base instructions that come from TrustYou. Every Agent runs on a core prompt we write and maintain, and it's always active on every message, no matter what the guest asks. It's what makes the Agent behave like a careful, grounded hospitality assistant out of the box. Your Custom Instructions are layered on top of it (more on that below), so you're adding to that foundation, not replacing it.

What the Agent Can't Do

Understanding the limits is just as important as understanding the capabilities:

No internet access. The Agent can't browse the web, search Google, or fetch information from external websites. It only works with the knowledge you've provided.
No memory across conversations. Each conversation starts fresh. The Agent doesn't remember past chats with the same guest (unless that history is available through guest context from a connected source).
No self-configuration. The Agent can't enable or disable its own Features, add knowledge, or change its settings. You control all of that from the backoffice.
Grounded in your knowledge. The Agent is designed to base its answers on the knowledge you've provided. While AI responses can occasionally be imprecise, the Agent prioritizes accuracy and will typically defer to a human rather than speculate.

Custom Instructions: Use Them Sparingly

Custom Instructions let you add your own guidance on top of the Agent's built-in behavior. They're useful, but they're a sharp tool, and they are not the place to solve most problems. For anything factual, the answer is almost always knowledge, not instructions.

Instructions are always on; knowledge is looked up

This is the key difference. Your knowledge is searched per question: the Agent only pulls in the pieces relevant to what the guest asked. Your Custom Instructions are different, they're injected into every single message, no matter what the guest asks. They're always part of the Agent's core prompt.

That's why a little goes a long way. The more you put in Custom Instructions, the more text sits in front of the Agent on every message, competing for its attention with the guest's actual question and the knowledge it just looked up. Pile on too much and the Agent gets defocused: its answers drift, miss the point, or follow your standing guidance when they should be following the guest. A short, sharp set of instructions keeps the Agent focused; a long one pulls it in every direction at once.

Why heavy instructions are risky

Custom Instructions modify the system prompt that drives the Agent. We improve that underlying prompt over time. If you've fine-tuned long instructions to chase a specific goal or patch particular edge cases, those tweaks are built against today's prompt, and an update on our side can make them behave in unexpected ways. Lightweight, general instructions survive updates gracefully. Heavily engineered ones don't.

What to use them for

Keep Custom Instructions short and general. Good uses:

Tone of voice hints, beyond the built-in Casual/Formal setting (for example, "warm and concise, never pushy").
A few specialities about your hotel worth emphasizing (for example, "we're a family-run, pet-friendly property; lead with that warmth").

For example, a backpacker hostel for young surfers might write:

Keep it relaxed and friendly, like talking to a fellow traveler: use first names, skip the formality, and a little surf slang is welcome. We're a budget spot built around the beach and the social scene, so lead with that energy and point people to the waves, board storage, and good cheap eats nearby.

That's the whole thing, two sentences. It sets a clear tone and a couple of specialities without scripting answers or listing facts.

Avoid using them to script answers, handle edge cases, or encode facts. Facts (including what you don't offer) belong in your knowledge, where the Agent looks them up when they're relevant. See Knowledge Best Practices.

Want a specific answer to a specific question? That's a job for knowledge, not instructions. Add the question and the answer you want as an FAQ entry in a Resource (a Markdown, Document, or Website Resource works well). When a guest asks something close to it, the Agent finds that entry in its knowledge and answers from it, exactly the wording you intended, without cluttering the instructions that run on every single message.

Custom Instructions don't change what the Agent knows. They nudge how it uses what it knows. Think of them as a couple of coaching notes, not a rulebook.

Putting It All Together

Here's the full picture:

You create Places and fill them with knowledge (Resources)
You create an Agent and assign it to the right Places
You enable Features to give the Agent capabilities (handover, quick actions, guest context)
You write custom instructions to shape how the Agent communicates
You connect Channels (website, WhatsApp) so guests can reach the Agent
Guests start chatting. The Agent searches its knowledge, uses its enabled Features, and responds within the boundaries you've set.

The Agent is the bridge between the knowledge you've curated and the guest asking a question. You control what goes in. The Agent handles how it comes out.

What's Next

Creating Your First Agent -- set up your first Agent step by step
Understanding Your Agent's Knowledge: Places and Resources -- go deeper on the knowledge side of the model
Adding a Website Resource -- start building knowledge from your website
Adding Document and Markdown Resources -- add knowledge from files or write it directly