Field notes on research agents
A demo from a company name.

I want to share one of my early experiences building research agents. It feels like a long, long time ago. LangChain had released LangGraph, and I was just getting into building agents.
Arick Vigas Autumn 2025
One-minute briefHere’s the task.
I was a contractor on a job for a UK company. Their main product was a tool for running deep analytics on a business’s database. They had an army of salespeople, and they wanted to make them even more efficient.
A salesperson would spend about a week preparing a demo for a client. And it was still dry. I mean, slides, features, examples from other companies…
They wanted the salesperson to show a demo using exactly the kind of data the customer might have. Imagine you want to show your analytics tool to a coffee chain called Strbkcs. To model by hand what their database, KPIs and data anomalies might look like… well, the preparation could take a long, long time.
My ambitious client wanted to automate it. The salesperson just types the name of the company into this magic tool. And…
Well, what would you do with that task?
Right. We’d Google the company.
So that’s the first step for the agent. A bit of simple code to normalize the link or company name, then some searches and crawlers to find the relevant pages.
Then we need to normalize all the data we’ve got. Some of it is Markdown, some is just HTML. Some companies have such deep websites and so many departments that you really want to set some boundaries. Otherwise you start mixing up departments.
Some companies didn’t like crawlers at all…
I decided to use Firecrawl at that point. Spending a lot of time on an almost-solved problem made no sense to me. Better to focus.
Okay, we have the basic research. What’s next?
Well… to make a good guess about what database a company might have, and what would be worth analyzing, what do we need?
Every company worships KPIs, right? What you don’t measure, you can’t control. Or how does the saying go?
So we put the research together and give it to an LLM. We ask for structured output, with fields like industry, subindustry and so on.
Then we give the model the research, the industry and the other details for another pass. This time we ask it to suggest KPIs. Better to ask for the five main ones first, then the secondary ones.
What if it hallucinates? Let’s put a human at this step to check and correct the suggestions. Fine.
What’s next?
Well, now that we have all this data, let’s give it to the LLM again with another prompt. This time we ask it to generate a database schema using a familiar approach, like a star schema.
Again, let a human approve it and make corrections if needed. And of course, we save each step in the database!
Then we take our data lake and let the fun begin. In wild loops, we generate thousands of records and link them together. That’s a tricky part. I won’t give away all the secrets, but ask your favorite agent, or do a little more research, and I bet you’ll find a nice way to do it.
I’ll share something better here. A few hidden problems to watch out for.
100,000linked synthetic records.
First, localization.
If your company is Japanese, generating a bunch of John Does won’t look so good. And if you add localization for a company that uses a right-to-left script… well, be ready for right-to-left text :D
Then there’s memory.
Once you’re generating hundreds of thousands of records and linking them together, make sure your process won’t run out of memory. And if a hallucination breaks the logic that links the records, the data will be really hard to fix at that stage.
Add some variation.
While all this is generating, you want the data to take seasonality and the specifics of the industry into account. Plan those variations ahead of time!
Almost ready.
Generate the charts. Then take the interesting points you’ve built into the data and prepare a guide for the salesperson. What should they show in Snowflake? How can they use it to deliver a great pitch?
Frankly speaking, a lot of this process was already designed when I got my hands on it. I helped where the pieces joined up, worked on some of the logic and made improvements along the way.
So what we had was basically a ten-step state machine. A linear graph, with some steps that used a mix of LLMs and APIs.
One foot in R&D and the other in business.
The thing is, when you’re building something like this and you haven’t seen anyone build it before, you’ve got one foot in R&D and the other in business. Better to make things configurable and versioned right away.
We versioned every prompt and kept control of each revision. That way we could change one thing and see what it affected. It’s much simpler than redeploying code. Trust me.
Now I think you could just ask your agent to implement this plan. Or just /goal it.
But hey… what if you want to save some tokens and have a little determinism?
Maybe an agent will read this article after all, and learn something…