Some of the most meaningful improvements at Yolando start the same way: with someone noticing a tiny friction point and caring enough to look closer. What begins as a small question often becomes a real contribution—sometimes even beyond our own product.

That’s what happened when our Founding Engineer, Sarth Frey, spotted a tiny inconsistency deep in our onboarding workflow. It wasn’t causing chaos. It wasn’t on a checklist. But it was just irritating enough that he couldn’t ignore it, and decided to take a closer look.

That instinct, to pause and investigate something small, was what ultimately led us to uncover a reliability issue that had been quietly affecting every new customer who joined the platform. 

A Small Reliability Issue With Outsized Consequences

When new customers join Yolando, they start by sharing their website URL. From that single input, our system has to form a clear picture of who they are—their products, personas, competitors, and overall positioning. 

“All of these different signals inform what we do in the rest of the app… it’s really important during onboarding to gather this context ahead of time and truly understand it.”

To make that happen, Pydantic AI sits at the core of our agent framework, coordinating the LLM calls and built-in tools we use to interpret a customer’s website and generate the structured signals that shape their onboarding experience.

It’s what powers tasks like fetching raw site data through Google’s URL context tool—the first step in understanding a brand when all we have is their homepage.

That early summary powers everything that follows, which means the LLM call behind it has to be reliable. But over time, the team noticed something subtle: the Google URL context tool—used through Pydantic AI—was returning empty metadata about 10% of the time. In those moments, the model had nothing to analyze. No content meant no summary, and onboarding simply couldn’t move forward.

As Sarth explained,

“Everything else we need to determine during onboarding is contingent on that first step. Knowing what the brand does is at the core of everything.”

Retries didn’t fix it. Some URLs returned null data multiple times in a row. It wasn’t random, it was a limitation baked into the tool.

Most teams would have patched around it. Ours didn’t. 

A Better Alternative—Just Out of Reach

Anthropic’s Claude already had a solution: its own built-in web fetch tool, capable of retrieving and summarizing website content in the same way Google’s tool does. If we could use Claude whenever Google failed, the entire onboarding flow would immediately become more reliable.

But there was a catch. Pydantic AI didn’t support Anthropic’s tool yet, it only supported Google’s. Because so much of our agent workflow depends on Pydantic AI, switching libraries wasn’t realistic. And accepting a 10% failure rate wasn’t acceptable either. So Sarth started digging.

“I did a quick search to see if anyone also had asked about adding this support for Anthropic’s web search tool to Pydantic AI. And I did actually find an issue, and conveniently it had been marked eight hours prior with a label that said ‘good first issue.”

It was a clear sign that the maintainers not only knew about the gap—they were open to someone new stepping in to solve it.

Looking back, he laughed about the timing.

“That was kind of unbelievable,” he said. “In retrospect, it’s crazy that I checked right after they marked it.”

And with that timing lined up, he decided to take it on before someone else did.

His contribution ultimately became PR #3427 in the Pydantic AI repository, adding full support for Anthropic’s web-fetch tool and improving reliability for everyone using the framework.

The first version came together quickly—just a day or two to get something working. But turning that into production-ready code was the real lift.

As Sarth put it,

“In Engineering, you spend 20% of your time getting the thing working, and 80% making sure everything affected by it isn’t broken.”

After a round of reviews and refinements, the PR was merged. The update is now part of Pydantic AI’s core toolkit, and onboarding at Yolando became noticeably more reliable almost immediately. 

Why This Story Matters More Than the Code Change

Yes, the contribution improved our onboarding. Yes, it made our workflow more stable. And yes, it solved a real problem for other teams using Pydantic AI. But the deeper significance of this moment isn’t technical at all. It’s cultural.

At Yolando, curiosity isn’t a side trait, it’s part of our DNA. It’s how this company was built in the first place, and it’s still what drives us today. When something feels off, even in a small way, we look closer. We ask why. We fix what isn’t working, even if it lives outside our own codebase.

“I’ve been at companies where an engineer sees a problem and they’re told it’s out of scope. Here, if there’s an issue with a package we don’t maintain, maybe you should just go and fix it.”

That mindset — the instinct to step outside the boundaries, improve something foundational, and quietly make the ecosystem better — is exactly the type of engineering culture we’re building. Curious enough to trace a problem to its root. Confident enough to dive into complex open-source systems. Thoughtful enough to make changes that help more than just ourselves.

The PR itself was modest. The ripple effect wasn’t. 

A Quiet Improvement With Lasting Reach

What began as an annoying onboarding inconsistency turned into something much bigger: a meaningful contribution to a major AI library, a smoother experience for everyone who uses Yolando, and an example of what engineering at this company looks and feels like.

It didn’t require a six-month project or a major initiative. It required someone who cared enough to pause, look deeper, and make things better for everyone.

We’re grateful for the work, proud of the contribution, and even more proud of the culture it reflects.

And we’ll keep watching for the next small friction point—because around here, those moments tend to lead somewhere worthwhile.

If that sounds like the kind of engineering culture you want to be part of, reach out at info@yolando.com!