TL;DR
I built an AI research agent that actually browses the live web and finds tech events, no search loops, no retry logic, no hallucinations. Just ask a question and get structured JSON back with the reasoning steps included. The secret? An API that handles multi-step research automatically. Built with .NET/Blazor in a weekend. Watch the video | Get the code | Free API key
(version française)
Happy New Year! I wanted to share something I recently presented at the AI Agents Conference 2025: how to build intelligent research assistants that can search the live web and return structured, reliable results.
Coming back from the holidays, I'm reminded of a universal problem: information overload. Whether it's finding relevant tech conferences, catching up on industry news, or wading through piles of documentation that accumulated during time off, we all need tools that can quickly search and synthesize information for us. That's what Reka Research does, it's an agentic AI that browses the web (or your private documents), answers complex questions, and turns hours of research into minutes. I built a practical demo to show this in action: an Event Finder that searches the live internet for upcoming tech conferences.
The full presentation is available on YouTube if you want to follow along: How to Build Agentic Web Research Assistants
The Problem: Finding Events Isn't Just a Simple Search
Let me paint a picture. You want to find upcoming tech conferences about AI in your area. You need specific information: the event name, start and end dates, location, and most importantly, the registration URL.
A simple web search or basic LLM query falls short because:
- You might get outdated information
- The first search result rarely contains all required details
- You need to cross-reference multiple sources
- Without structure, the data is hard to use in an application
This is where Reka's Research API shines. It doesn't just search, it reasons through multiple steps, aggregates information, and returns structured, grounded results.
The Solution: Multi-Step Research That Actually Works
The core innovation here is multi-step grounding. Instead of making a single query and hoping for the best, the Research API acts like a diligent human researcher:
- It makes an initial search based on your query
- Checks what information is missing
- Performs additional targeted searches
- Aggregates and validates the data
- Returns a complete, structured response
As a developer, you simply send your question, and the API handles the complex iteration. No need to build your own search loops or retry logic.
How It Works: The Developer Experience
Here's what surprised me most: the simplicity. You define your data structure, ask a question, and the API handles all the complex research orchestration. No retry logic, no search loop management.
The key is structured output. Instead of parsing messy text, you tell the API exactly what JSON schema you want:
public class TechEvent
{
public string? Name { get; set; }
public DateTime? StartDate { get; set; }
public DateTime? EndDate { get; set; }
public string? City { get; set; }
public string? Country { get; set; }
public string? Url { get; set; }
}
Then you send your query with the schema, and it returns perfectly structured data every time. The API uses OpenAI-compatible format, so if you've worked with ChatGPT's API, this feels instantly familiar.
The real magic? You also get back the reasoning steps, the actual web searches it performed and how it arrived at the answer. Perfect for debugging and understanding the agent's thought process.
I walk through the complete implementation, including domain filtering, location-aware search, and handling the async research calls in the video. The full source code is on GitHub if you want to dive deeper.
Try It Yourself
The complete source code is on GitHub. Clone it, grab a free API key, and you'll have it running in under 5 minutes.
I'm curious what you'll build with this. Research agents that monitor news? Product comparison tools? Documentation synthesizers? The API works for any web research task. If you build something, tag me. I'd love to see it.
Happy New Year! 🎉
