At Core Automation, Jerry Tworek is trying to hand AI research over to agents. Much of the time, he said, they let him down.
“Whenever agents fail for me, I think, ‘Okay, we still have jobs, we still have work to do. There is a reason for us to be here,’” he said on Nebius’ AI-Native podcast, in an episode released Nov. 18.
Tworek is the founder and CEO of Core Automation. The lab wants agents to conduct research that produces better agents, which could then take on more of the research. He said the agents he uses still struggle with high-level research reasoning and software system design.
“I have a lot of trust issues with those models,” he said.
He spent nearly seven years at OpenAI, where he became a vice president of research. He left in January 2026 and later founded Core Automation. The company has since attracted hundreds of millions of dollars in investment.
Tworek wants his new team to push agents just beyond what they can reliably do.
“If you don’t interact with slop each day, probably you aren’t automating hard enough,” - he said, citing one of the company values he wrote.
He holds agents to the standard of any new hire: do the job, learn something, do it better next time.
“Can we see day over day our models being more useful for our work?” he said on AI-Native.
An education for the machines
Core Automation is pursuing continual learning: models that keep learning from new experience after their initial training. Tworek compared the goal to the difference between a junior researcher and a veteran whose past work changes how they approach the next problem.

Jerry Tworek (left), founder and CEO of Core Automation, speaks with AI-Native podcast hosts Markiesha Patrice (center) and Josh Liss (right). Credit: Nebius’ AI-Native podcast..
He wants models to carry those lessons forward. Patching one weakness after another, in his account, amounts to “playing whack-a-mole.”
The lab would be its own test case. If its agents became substantially more useful to its researchers, Tworek said, other companies would benefit, too.
The lab is also exploring new learning methods and alternatives to the architecture used by many current AI models. Tworek expects surprising results from changing the types of models being trained, rather than simply making existing designs larger.
A model that keeps learning could also learn to comply with requests it was trained to refuse. His approach: start with a few enterprise customers he can work with closely. “Companies have their own liability and their own responsibility towards the world,” he said. He expects those obligations to discourage misuse.
Working with a few customers would let researchers observe how the systems develop and build safeguards before a wider rollout. Asked whether he was confident AI would be controllable, he said: “I would like to have that confidence. I believe that it is possible.”
Freedom, with responsibilities
Tworek’s preferred future is one in which almost anyone has a team of agents working on their behalf. “Everyone should feel like they are the CEO of a company,” he said. People would still monitor the agents and make sure they were spending resources appropriately. The CEO job comes with management duties.
Tworek hopes AI will take over enough labor to give people more control over their time. He described coming home at 10 p.m. and wanting more time with his three young daughters. He also asked what would replace work in people’s lives. For many, he said, it provides a reason to get out of bed and attempt something difficult. “AI is here to give us freedom, but what do we do with that freedom?”
For now, the agents he uses still leave him with plenty to do.
“If the AI is so good and so smart, why do we have to come to work each day?” he asked. “Why is there still so much to do?”
Watch the AI-Native episode with Jerry Tworek:



