An August 10 essay from MIT Technology Review argues that autonomous agents, built to mirror how researchers actually work, will drive science’s next chapter. Raw data alone, it contends, will not get the field there.
The piece points to AlphaFold as both inspiration and warning. The protein-folding breakthrough showed what AI plus enough data can achieve, and it helped launch a wave of startups building foundation models for biology, chemistry, and materials that raised billions. But the essay contends that treating AlphaFold as the template for all of science misses what research actually is.
Science is a process of forming hypotheses, designing experiments, questioning results, and revising theories. Feeding a model more data does not reproduce that process. What could, the argument goes, are AI agents that model the human research workflow itself, proposing experiments, interpreting failures, and deciding what to try next.
The distinction matters for healthcare and biotech, where investors and executives are betting that AI can shorten drug discovery timelines. If the bottleneck is reasoning about experiments rather than pattern matching on data, then progress depends on agentic systems and evaluation methods that measure scientific judgment, not just prediction accuracy.
That view is gaining ground across the field, with labs and startups racing to build autonomous research agents. Whether they deliver on the promise will depend on how well they capture the messy, iterative core of scientific work.
