← Things People Say on Podcasts

“And I have seen this product being built at companies more than any other product. Like everyone is building this. It is some kind of agent, internal agent that will answer business questions. But this is actually really bad. Do not make something like this.” Hamel Husain · Vanishing Gradients

Hands-On Engineering Podcasts · September 2026

“And I have seen this product being built at companies more than any other product. Like everyone is building this. It is some kind of agent, internal agent that will answer business questions. But this is actually really bad. Do not make something like this.” — Hamel Husain, Vanishing Gradients

This is the clip Vanishing Gradients opens the episode on, and host Hugo Bowne-Anderson picks it straight up to ask why the design is bad. Husain's answer in the next turn is short: you cannot tell whether the agent's answer is right. He is describing the internal AI product he sees companies building more than any other.

Transcript

Vanishing Gradients Around 00:00 into the episode
Hamel Husain

And I have seen this product being built at companies more than any other product. Like everyone is building this. It is some kind of agent, internal agent that will answer business questions. But this is actually really bad. Do not make something like this.

Hugo Bowne-Anderson

Yo, so that was Hamil Hussein, ML engineer, longtime friend of the Vanishing Gradients podcast, and one of the most interesting, dare I say, AI scientists in the space. Why is the data agent everyone is building such a bad design?

Hamel Husain

You don't really know if it's right. How do you know?

Hugo Bowne-Anderson

Data science is back, baby. Every agent produces noisy outputs, tangled tool calls, strange human behavior, and thousands of possible paths through a system. Hamill and I trace what that means from the data agent everyone's building to a possible emerging new role, that of the AI scientist. We also explore agents helping to evaluate other agents, learning from human annotations and servicing failures that teach the reviewer something new. Along the way, we get into product design, verification, search, human judgment, and why the proof behind an AI-generated answer may be more valuable than the answer itself. Hamil and Shreyashanka's AI evals for Engineers and PMs course begins September 5th, and I'll be taking it for a fourth time, I think now. The link and a 25% vanishing gradients discount are in the show notes. We recorded this conversation live and we live stream many episodes and run free workshops. And you'll find our events calendar, subset newsletter, and YouTube channel in the show notes. If you enjoy the episode, please share it with someone building an AI product right now. I'm Hugo Bown Anderson and welcome back to Vanishing Gradients. Hey there, Hamil, and welcome back to the show.

Speaker names from our own diarization · position estimated from where the line sits in the episode