What Is an Agentic Website? Lessons from Onz in Singapore

An agentic website doesn't just show you pages. You give it a goal and it does the steps for you. On a normal site, you do the looking up, comparing and planning yourself. On an agentic one, the site takes your request, checks real data, makes the choices and hands back something finished.

We built one for ourselves. Onz is Rampage Connect's own events site for Singapore. Its main feature, My Day, takes a single line like "chill Saturday with my partner" and turns it into a day plan: real events, lunch at a hawker centre or a restaurant, and how long it actually takes to get between stops. The weather is live from NEA and the travel times come from OneMap. Say "halal" and every meal it suggests is MUIS-certified.

This is the story of why we built it and what "agentic" actually means once you're past the buzzword. It also covers the lessons a Singapore business can use for its own AI features. I'll stick to what anyone can see on the live site. Where I describe design choices, they're the reasons behind the build, not a diagram of how it's wired.

---

The problem: Singapore's weekends are planned in 12 tabs

Planning a Saturday in Singapore goes something like this. You scroll a couple of event listings, a museum's Instagram, a Telegram channel and whatever a friend forwarded. You find something at the Esplanade at 3pm. Then you open Google Maps to see if you can get from Tampines to City Hall in time. Then you check if it's going to pour. Then someone in the group chat asks where to eat, and whether it's halal, because one of you doesn't eat pork.

Every piece of information exists. It's just scattered, and joining it up is left to you.

When we looked at what already ranks for "AI itinerary planner Singapore", nearly all of it is built for tourists. You get multi-day trip templates: Marina Bay, Gardens by the Bay, Sentosa, repeat. Useful for a first visit, but none of them know what's actually on this Saturday, or whether the hawker centre you're being sent to has a certified halal stall.

That gap is the reason Onz exists. We weren't trying to build another trip planner. We wanted something that answers the question locals ask every week: what do we do this weekend?

---

What "agentic" means in practice (vs a chatbot bolted on)

Everyone is putting a chat bubble in the corner of their site right now. That isn't what makes a site agentic. The industry hasn't settled on a definition yet, but the test we use is simple:

Does the site do the work, or does it hand the work back to you?

A chatbot on an events site would tell you "there are many exciting events this weekend!" An agentic events site gives you a plan: this exhibition at 11am, this hawker centre for lunch, 25 minutes by MRT to the next stop, indoors in the afternoon because there's rain in the forecast.

The word "agentic" sounds grand, but the principle is modest. Take one tedious, multi-step task your users already do by hand, and do it for them properly.

---

How My Day works, from the user's side

Here's what you see on onz.sg today.

1. One line, not a form. The prompt says it plainly: "Don't scroll. Just say what kind of day you want." There's no stack of filters for date, district, budget and dietary needs. You write a sentence the way you'd text a friend.

2. Presets for people who don't want to type. There are ready-made starting points: a fun-filled day with children, a makan crawl in the East with friends, a chill Saturday with a romantic partner, and halal makan with family. These matter more than they look. Many people have never used a planner like this before, and seeing a few examples shows them what they can ask for.

3. A whole day comes back. The result has real events from Onz's listings, meals at a hawker centre or a restaurant, and the travel time between each stop. That last part is what makes it usable. A plan that ignores the 40 minutes between Jurong and Pasir Ris isn't a plan.

4. Halal is a hard rule, not a tone. If you say "halal", every meal is limited to MUIS-certified options: 739 halal restaurants, plus the certified stalls in 29 hawker centres. It won't guess from a menu, and "Muslim-friendly" on a signboard doesn't count.

5. It sits inside a normal events site. You can still browse My Day, This Weekend, Free, Family, Date Night and All Events. There's a weekly newsletter (the "weekly jio"), and organisers can submit their own events. The agentic feature is one entry point, not a replacement for browsing.

---

Why grounding in real data is the whole game

Most people take the wrong lesson from AI demos. The clever part isn't the language model writing a nice paragraph. The clever part is limiting what it's allowed to say to things that are true today.

If you leave a general model on its own, it will happily plan your Saturday around a festival that ended in July, a café that closed in 2024, and a "quick 15-minute hop" from Woodlands to Marina Bay. It sounds right and it's wrong. One bad plan and the user never comes back.

So My Day is grounded in three kinds of real data, each from the source that owns it:

There's a pattern here. For every fact a user would act on, we went to the authority for that fact rather than whatever a model "remembers". Weather comes from the weather authority, routes from the national map, and halal status from the body that certifies it.

That rule matters far beyond events, and it's the same thinking behind how we approach AI in marketing generally. If your AI feature quotes prices, stock, availability or eligibility, those answers must come from your live system, not from the model.

---

Where Jev fits: fast decisions for curation

Planning is only half the job. The other half is curation: taking a constant flow of events from public sources and organiser submissions and sorting them. Which category is this? Is it really family-friendly? Is it free? Does it belong in Date Night?

Those are small decisions, and there are a lot of them. For that part of Onz, curation runs on Jev, a decision model from TypeSafe AI. Jev doesn't write text; it answers structured questions — pick one of these, rate this, is this true — and says how confident it is. Jevable, a showcase of community projects, gives a sense of what people are doing with it.

The business lesson is bigger than any one model: use AI for clear, structured decisions, and act on its confidence. Let high-confidence answers through, review the middle band, and send low-confidence ones to a person. That's the right way to run any AI decision in a business, whichever model sits underneath.

---

Lessons for your own business's AI features

You probably don't need an events planner. But the same thinking applies whether you run a renovation firm, a tuition centre, a clinic or an F&B group.

1. Pick one tedious, multi-step task

Don't start with "add AI". Start with the job your customers find annoying. A renovation lead works out a rough scope for their HDB flat. A parent compares class timings across three branches. A diner wonders if you can fit 12 people on a Friday. If customers currently need four WhatsApp messages to get an answer, that's your candidate.

2. Only let the AI choose from data you actually maintain

This is the rule that makes or breaks it. If your service list, availability or prices live in someone's head or an outdated PDF, fix that first. An agent is only as good as the most out-of-date data it can reach.

3. Use typed decisions wherever you can

Routing an enquiry, tagging a lead, deciding which package fits: these are choices, not essays. Structured outputs are faster, cheaper and far easier to check than free text.

4. Build in "I'm not sure"

Set confidence thresholds. When the system isn't confident, it should hand over to a person. In Singapore that usually means WhatsApp. A graceful handover beats a confident wrong answer every time.

5. Hard rules beat soft vibes

On Onz, halal is a filter, not a suggestion. Find your equivalent. Maybe it's "never quote below this floor", "never book outside operating hours" or "never promise a date the roster can't cover". Put those rules in code, not in a prompt.

6. Keep the normal website working

My Day sits alongside ordinary browsing. Some people want to be guided and some want to scroll. Don't force everyone through the agent.

7. Mind PDPA from day one

If your agent collects names, phone numbers or preferences, the PDPA applies: get consent, collect only what the task needs, and be clear about how you use it. Onz's newsletter sign-up says what the email is used for, in one line. Aim for that level of clarity.

---

Mistakes to avoid

• ▸Calling a chat widget "agentic". Your customers will notice when it can't actually do anything.
• ▸Launching without a data owner. Someone must be responsible for keeping the source data current, or the feature decays quietly.
• ▸Letting the model invent specifics. Prices, dates, addresses and certifications must come from records, never from what the model "remembers".
• ▸Hiding the fallback. Always make the human route obvious.
• ▸Measuring the demo, not the repeat use. The real signal is whether people come back and use it a second time.

---

If you're thinking about an AI feature of your own

Onz is our own product, so we get to experiment on it, and the case study at /work/onz-singapore goes into more detail. The lessons carry over to client work: pick a narrow task, ground it in data you trust, return typed decisions where you can, and always leave a way to reach a person.

If there's a task on your site that customers find tedious, have a look at how we approach AI agents and automation. Or just drop us a WhatsApp and tell us what you're trying to do. We'll tell you honestly whether an agent is worth building, or whether a better form and a faster reply would do the job.