Some Things in Life Just Have to Happen. AI Still Can’t Help With the Hardest One.
You can’t marry someone over API and your kids need to physically be in a school.
There are things that have to happen in person
You can’t marry someone over API. Your kids need to physically be in a school. If your company needs you in SF because that’s where the talent is, you actually have to live in SF. No amount of technology removes the fact that some of life’s biggest moments require you to be somewhere, and getting there is one of the most stressful things you’ll ever do.
AI just crossed a real threshold. Matt Shumer's viral post Something Big Is Happening described walking away from his computer for four hours and coming back to finished software, and that tracks with what we're seeing too. The models are genuinely that capable now. And at the same time, Anthropic's head of safeguards research just resigned, warning that the world is "in peril". The capability is real and accelerating, and the stakes of getting it wrong in the real world are higher than most people in AI are grappling with.
Relocation sits at the intersection of everything that makes AI hard: long timelines, cascading dependencies, fragile and irreversible decisions, messy real-world inputs, and deeply personal stakes. It’s not just a logistics problem. It’s immigration law, housing markets, school systems, pet regulations, and payroll all moving at once, all depending on each other, with a real family in the middle of it.
The engineers working on this are solving one of the most complex and consequential operations problems that exists, because when it goes wrong, the repercussions aren’t abstract. Someone’s visa gets delayed, their kids can’t start school, their belongings are stuck in a port.
This is what we are solving for at Gullie, and here’s what we’ve actually learned.
Memory is everything, and almost nobody has it figured out.
A relocation unfolds over months. Someone casually mentions they’re bringing their dog, and three weeks later that matters because of quarantine rules in the destination country. If your agent dropped that detail, you end up with a pet stuck at the border.
About a year ago we started treating every conversation as a database, converting each interaction into structured, persistent information that includes entities, relationships, timelines, and dependencies, and that updates as the situation changes.
The industry calls this "context engineering." We were just calling it the only way any of this works. It's catching on. Mastra just shipped their Observational Memory (OM) system, and Letta has been doing interesting work on agent memory architectures.
One sentence can break six things
Someone says “we’re pushing the start date back two weeks.” It sounds simple, but it means immigration timelines shift, which means temporary housing dates change, which means the lease start moves, which means shipping might not work anymore, which means storage needs rearranging.
Most agent architectures are built to handle individual tasks, not webs of dependencies that ripple across each other. Figuring out how to propagate changes like this reliably is the actual hard engineering problem, and it’s the one we spend the most time on.
You can’t move fast and break things with someone’s passport
This is the part that separates what we’re building from everything else in the agent space right now. When you’re dealing with immigration paperwork, bank accounts, visa timelines, and lease agreements, you’re operating in a world where a single mistake can have real and sometimes irreversible consequences.
OpenClaw went viral because people loved the feeling of AI that “does things,” and then the security problems started surfacing. We’ve seen firsthand what happens when agents get too much autonomy with sensitive information. The answer isn’t to avoid autonomy entirely, it’s to earn it incrementally. We expand what our agents handle one workflow at a time, proving reliability before extending trust. And some things just need a human in the loop, maybe always.
The real world communicates in human slop
There’s AI slop, and there’s human slop.
Vendors don’t send structured data. They send PDFs, email threads, spreadsheets, portal screenshots, and updates that directly contradict what they said last week. In shipping, a surprising amount of work is still done on paper.
An agent can reason about a messy email, pull out what matters, and flag when something doesn’t add up, which gives it a real advantage over traditional automation that just breaks when the format changes.
But a misread shipping date becomes a missed delivery window, which becomes a family sleeping on an air mattress for two weeks. The stakes are high because someone has to live with the inconveniences of any mistakes being made. So proper tracking, logs and documentation are crucial.
The only question that matters
We don’t evaluate our agents by asking “did the model answer correctly?” We ask whether the person’s move actually went well, and whether they felt like someone had their back during one of the most important transitions of their life.
People move to get married, to start a new job, to be closer to family, to give their kids a better school. These are the moments that shape how your life turns out, and the way we think about building for them has to be different.
Every dependency matters. Every detail matters. The margin for error is close to zero, and the human on the other end is going through one of the most stressful experiences of their life.
That’s the problem we’re working on.
We’re looking for people who want to help make great relocation support affordable and robust for everyone. Because right now, getting world-class help during one of the most important transitions of your life is something only a few people have access to, and we think that should change.
So, if you’re an engineer who wants to work on something that’s both technically challenging and genuinely matters in people’s lives, come join us.