Bangladeshi engineer making AI agents reliable

When Andrej Karpathy, a founding member of OpenAI, sat down on the Dwarkesh Podcast and the current generation of AI agents "slop" that would take "about a decade" to actually work, he was saying out loud what a lot of engineers building on top of agents had been quietly admitting in private.

Mostofa Adib Shakib is one of them. The Bangladeshi engineer who is now building Variant Labs, an early-stage company betting that the gap between what AI agents demo and what they ship to customers will not close on its own.

"You'd be on a call with a founder," Shakib said, "and they'd show you this incredible agent demo. And two weeks later, you'd be back on a call and they'd be telling you it broke for their first paying customer. That kept happening."

Agents hallucinate. They pick the wrong tool. They cannot recover when one step fails. Gartner is forecasting that companies will scrap more than 40% of agentic AI projects by 2027, and recent benchmarks of frontier models put their first-attempt success rate on real-world tasks closer to a quarter than to anything resembling production reliability. Variant runs agents through realistic scenarios, surfaces the failures, and turns those failures into data that data engineers can use to make the next version less likely to break the same way.

It is a long odds bet. The category Shakib has chosen is already crowded. Braintrust, a competitor attacking a similar problem, raised an $80 million Series B this year at an $800 million valuation, with backing from Andreessen Horowitz, Greylock, and Elad Gil. Patronus AI,

LangSmith, and Arize are all well capitalised and years ahead in market presence. Above them, some of the loudest voices in the industry are betting on the opposite thesis altogether, that the next generation of models will be smart enough to make a reliability layer unnecessary. Either Shakib is right and he is one of the smallest players in a category that will define how AI gets deployed, or he is wrong and Variant is solving a problem that disappears.

He seems unbothered by the math. "There's a lot of capital moving for stories right now," he said. "I want a real product, real customers, a real problem. The story can come after."

What gives the bet some weight is where Shakib spent time before Variant. At Snapchat, he led the identity migration that let users change their usernames, a system that touches roughly a billion accounts. Identity at Snapchat is the foundation every other part of the product depends on. Every chat. Every friend graphs. Every login. Every integration. A bug in a migration at that scale locks real people out of their accounts and cascades through every system the user graph touches. The work is one of the more demanding reliability problems in consumer software.

Before settling on what to build at Variant, he spent months talking with researchers at frontier AI labs and engineers deploying agents in production. He wanted to see the failure modes from both ends, he said, the model side and "the inside of an on-call rotation."

Variant is not the first time Shakib has built infrastructure underneath an emerging wave. He earlier cofounded Promethium, a blockchain wallet and inheritance startup that received backing from the US government's SBIR program and the 1517 Fund. He also maintains Forge, an open source project that aims to generate synthetic data for training AI agents through reinforcement learning environments. "It's a pattern," he said. "I get pulled toward the parts of a system nobody's looking at. The boring parts and once I see it I have a hard time leaving it alone."

Shakib grew up in Dhaka and completed his undergraduate degree at Texas Tech University before joining Snapchat. He had also built a parallel career as a content creator for students back home, running YouTube and Facebook channels on programming, AI, and career strategy that crossed two million views and a combined following of around sixty thousand before he stepped back from posting to focus on building.

The messages still arrive. They tend to be variations of one question. How do I get in?

"Most of the smart kids I meet back home are still asking permission," he said. "Permission from their parents, permission from a senior, permission from a recruiter at a multinational. Nobody is going to give you permission to build the thing you want to build. You have to start, and then you ask later." That is, in effect, the bet he is making with Variant too. An unfashionable thesis. A wager that the giants currently betting against him will eventually need what he is building.