The line between a lab coat and a developer's hoodie is vanishing, thanks to AI agents. Omar Sanseviero from Google DeepMind recently laid out how much of what passes for "research" today—especially the tedious, empirical kind—is quickly becoming an engineering problem solvable by AI itself. For ambitious founders and builders in their 20s and 30s, this isn't just a trend in big tech; it's a blueprint for accelerating discovery and product development in your own company.

Key Takeaways

  • Research now often means "engineering with unknowns": Omar Sanseviero from Google DeepMind notes that many researchers spend their days on "ablations," which means moving pieces around to see what works. He believes this is "much more engineering rather than for like research," unless you are designing new fundamental architectures.
  • AI agents automate the grunt work of experimentation: DeepMind research teams are already building agentic tools to conduct experiments, run ablations, and perform evaluations. This frees human researchers from the empirical treadmill.
  • Beyond speed, agents discover novel paths: Swyx points out that agentic coding goes beyond merely "speed running experiments" you would have done anyway. The true game-changer is when agents explore and find “trajectories that people wouldn't think about” leading to genuinely new discoveries.
  • Your best researcher might be an agent: The shift means human researchers become less like manual testers and more like high-level directors, empowered by AI tools to pursue deeper, conceptual work, or interpret results from agent-driven exploration.

The Blurring Line Between Lab and Codebase

Forget the image of the lone genius toiling over equations. In the world of AI, much of what's called research looks an awful lot like engineering. Omar Sanseviero, a key voice behind Google DeepMind's Gemma 4, put it plainly: “so many researchers are doing ablations, right? Like they are just moving the pieces around and seeing what works and what doesn't work.” He sees this as distinct from fundamental architecture design, placing it firmly in the engineering camp. It's about iteration, testing, and optimization—tasks that AI agents are now swallowing whole.

This isn't theory; it's happening inside DeepMind. Sanseviero confirmed that “within the team we are building a skills to do experiments and ablations and evaluations.” Researchers are adopting these agentic tools not as an auxiliary, but “as part of their research process.” Imagine having an autonomous co-pilot that can run a thousand variations of your experiment while you grab coffee. That's the new reality.

Agents: Beyond Automation, Towards Discovery

Swyx drove this point home, describing the initial benefit as simply “speed running experiments agentically.” The agent, he says, is “more autonomous you can actually go to sleep and it will do the things that you would have done anyway.” This alone is a massive productivity bump. But the real shift is deeper than just doing the same work faster. It’s about doing different work—work human minds might never conceive.

"The side," Swyx continued, is when these agents aren't just shooting off predictable paths, but uncovering "trajectories that people wouldn't think about and they work and you make new discoveries." This moves AI beyond a tool for known tasks and into a true partner in discovery. It means the empirical search space, once limited by human intuition and stamina, now expands exponentially, governed by an AI with different biases and blind spots—ones that might just stumble upon the next big thing.

What to Do With This

Identify the empirical "ablations" in your own startup's workflow. This could be A/B testing UI variations, optimizing marketing copy, or iterating on a backend algorithm. Instead of tasking a human, build or buy a simple agent to run these iterative tests. Your goal isn't just to do it faster, but to free up your sharpest minds to focus on truly novel architectural challenges or product vision, knowing an agent might uncover an unexpected, superior path you’d never manually explore.