AI Agent Solves Decade-Long Superbug Mystery in 2 Days

José Penadés thought someone had hacked his computer when he saw what the AI had done.

The Imperial College London microbiologist and his team had spent nearly a decade figuring out how a family of superbugs spreads antibiotic resistance between species. In 2024, he described the unsolved problem to Co-Scientist, a Google DeepMind AI agent. Within two days, it returned five potential explanations ranked by likelihood.

The top answer matched exactly what his team had painstakingly proven over those long years. Some superbugs acquire tails from viruses and use them like keys to jump between host species. Penadés was so stunned he emailed Google to verify his computer hadn't been compromised.

It hadn't. The AI had genuinely worked it out on its own.

This moment reveals something profound happening in science right now. AI agents are different from chatbots because they can plan, break problems into steps, run multiple processes simultaneously, and catch their own errors. They can access databases, write code, and operate robots to pursue scientific goals.

The timing matters. Researchers face a growing "burden of knowledge" where the amount they need to learn before contributing meaningfully keeps expanding. Questions in drug discovery, climate modeling, and biology have become too complex for human teams alone to tackle quickly.

The Ripple Effect

What happened in software engineering over the past year is now happening in science. Coding agents reshaped how engineers work in barely 12 months. Science is messier but also well-suited for agents in certain ways. A massive literature exists as text, huge databases are available, and many workflows already rely on code.

Three forces make today's agents genuinely useful. Leading AI models now outpace human experts on demanding scientific knowledge tests. "Scaffolding" gives these models structure, memory, and tool access. Customization lets scientists tailor agents to their specific disciplines and workflows.

The challenge ahead isn't about whether agents will transform research. They already are. The urgent question is how to validate the coming wave of AI-generated ideas, ensure scientists can access these tools, and adapt peer review systems for this new reality.

Some of this change is perhaps overdue. Science depends on infrastructure that evolved gradually over centuries: labs, teams, peer review, grant funding. The era of AI agents will challenge that infrastructure and demand rapid adaptation.

For Penadés, one thing is clear: if he could have gone back in time with this insight, his team would have saved years of work.