# How AI Agents are transforming scientific discovery - Google DeepMind

*Источник: Google DeepMind*
*Дата: 2026-07-15*
*Язык: en*

**Кратко:** To some extent, agents simply add intensity to questions that policymakers are already focused on in their AI for Science strategies, such as how to train the next generation of scientists, how to experiment with new forms of scientific institutions, and how to ensure that AI is not misused by threat actors, while still putting the technology to use addressing the various natural risks that society faces, like the next pandemic. For example, AI agents could make it more feasible for small, agile teams to pursue creative, ambitious ideas, reversing the trend towards “big team science”, or enable scientists to work across domains, bringing new perspectives to existing problems.

To some extent, agents simply add intensity to questions that policymakers are already focused on in their AI for Science strategies, such as how to train the next generation of scientists, how to experiment with new forms of scientific institutions, and how to ensure that AI is not misused by threat actors, while still putting the technology to use addressing the various natural risks that society faces, like the next pandemic.
For every encouraging scenario, there is a challenging one. For example, AI agents could make it more feasible for small, agile teams to pursue creative, ambitious ideas, reversing the trend towards “big team science”, or enable scientists to work across domains, bringing new perspectives to existing problems. Assuming efficiency gains make agents sufficiently cost-effective, these trends could particularly benefit smaller, less well-resourced countries and institutions.
But agents will also give rise to anxiety among junior scientists that their institutions are choosing to spend budgets on tokens instead of staff. And left unmanaged, there is a risk that agentic tools could de-skill new generations of scientists before they develop the judgement needed to use them effectively. For the same reason that mathematics students still prove theorems unaided, science-graduate training may need structured periods of agent-free work and access to agents that act as genuine cognitive partners rather than oracles.
Beyond these questions, AI agents present at least four urgent new priorities: 1. Scientists need access to the tools. 2. The tools need access to agent-ready data. 3. We need more experimental infrastructure to validate AI ideas. 4. And we need to update the peer review process.
1. Ensure widespread access to agents
Agents will be extremely useful and fallible in non-obvious ways. Both factors provide a strong rationale for policymakers to ensure that all scientists can access the best agents — to speed up discovery and to provide the independent evaluations of AI agents that the scientific community needs to judge how best to use them.
This is an urgent strategic priority for policymakers and science funders, akin to the historical challenge of providing access to supercomputers. Geopolitical debates today often focus on one aspect of sovereign capability — whether a state can train its own frontier model. Much less attention is paid to what may prove to be a more important issue: a country's ability to deploy agents across its scientific ecosystems for transformative impact.
At the micro level, funders must first decide how labs and researchers select and pay for agents. Selection is the easier near-term problem: let researchers find the most useful tools for themselves, without excessive approvals or complex procurement.
Paying is harder. The temptation is to use existing structures, with scientists seeking funding through grant applications or drawing on lab budgets. But the compute required to run agents can be large and the frontier of what agents can do is constantly expanding. While the cost per unit of AI capability is falling fast, total lab expenditures on agents are still likely to rise as agents take on longer and more complex tasks. Policymakers must quickly assess whether budget uplifts or entirely new funding programmes are needed. Delivering access at the scale and price needed will require novel public-private partnerships; the US Genesis Mission is one promising model.
Bigger questions await. Agents may make the existing process of allocating national budgets across disciplines more legible, forcing science funders and research programmes to quantify the investment in data and compute needed to make progress on specific problems. This in turn may lead to more targeted debates about the relative value of solving different problems. If agents propose the top hypotheses to explore across an entire field, with very expensive experimental validation plans, how should this fit into national funding strategies?
2. Make national data assets agent-ready
While scientists need access to agents, agents need access to data. Data that is open or low-risk should be exposed to agents through well-documented APIs, with sufficient quality control and metadata. The engineering support and maintenance to do so is not trivial, so funders should ensure such data stewardship is properly resourced and support interoperable data standards. But ultimately, the ability of agents to help extract and annotate data — from PDFs to download portals — provides an opportunity for governments to extract a lot more value from the data they have already funded.
More sensitive datasets in genomics, virology, or other areas carrying dual-use risk often come with restrictions on who may use them and how. The challenge now is to develop similar privacy-preserving solutions, when appropriate, for agents, with auditability and privacy built in. Examples like OpenSAFELY, which lets human researchers securely access valuable health data, can provide inspiration. The prize is large. A dataset analysis that currently takes years of doctoral work could, with the right secure agent infrastructure, run autonomously in days.
Perhaps most importantly, agents provide a strong rationale for funding the creation of entirely new open datasets. This leads to a further question for funders: could agents help identify the most important datasets to fund? Some of the authors of this article recently made a human-expert-driven attempt to answer that question for fusion energy. How soon will agents be capable of running similar “AI data stocktake” exercises?
3. Tackle the validation bottleneck
Many scientists already struggle to get enough time in facilities to run their experiments. As AI agents make hypotheses and candidate solutions increasingly abundant, this bottleneck will only tighten. Policymakers and funders should address this in at least two ways: investing in existing experimental validation infrastructure and accelerating progress on automated labs.
Public research bodies hold extensive experimental facilities across almost every scientific field. AI agents provide a reinvigorated case for investing in them and opening them up, by renting bench space or experimental run-time to researchers testing computational hypotheses and predictions against reality. Direct partnerships with AI labs are another avenue. Google DeepMind has created a wet lab inside the UK’s Francis Crick Institute, a leader in biomedical research, and is also providing independent scientists funding — alongside Co-Scientist access — to carry out the wet lab experiments needed to validate agent-enabled hypotheses. The US government’s Genesis Mission will connect the world-class experimental facilities of the Department of Energy’s (DOE) National Laboratories with academia and the AI industry.
Automated labs are another promising route to tackling the validation bottleneck, but they currently rely on expensive robotics and compute. Ensuring broad access will require public investment, and there are good early efforts here. The US National Science Foundation has put $100 million towards a national network of distributed facilities, while the UK launched a call for ideas and already hosts the £81 million Materials Innovation Factory. The build-out of automated labs will almost certainly go beyond what any single lab or institution can afford, so governments should also explore building centralised capacity and adopting the kind of “user facility” access model seen at the US DOE’s national labs.
4. Empower peer reviewers with agents
The peer review process has long been under strain, with slow timelines and reviews of varied quality. Now, scientists are using AI to write ever more grant applications and papers. This is making it harder for funders to know which research to fund, and for peer reviewers to validate findings and identify the most important work.
As noted by Professors James Wilsdon and Geraint Rees, the challenge is not just an increase in supply; writing quality is also no longer a reliable discriminator. Agents deepen the problem. The more an agent is left to plan and optimise an application, drawing on the funder's criteria and its recent winners, the less the bid reflects a scientist's original thinking.
Funders, journals, and conferences are trialling various responses. For grant applications, some suggest leaning less on the written word and more on the investigator's track record and team. The UK's Medical Research Council recently reinstated interviews for shortlisted applicants. These ideas have promise, but also risk privileging seasoned experts or increasing costs.
The likely answer will be a layered approach. Those developing and using AI should document its use more clearly. That could include advancing watermarking techniques and ideas such as “Human-AI Interaction Cards" — short records detailing the prompts and outputs that produced the key scientific insights.
Reviewers should also be able to use agents. This will not be easy, as AI use is often banned, even if widely practised in secret. To move forward, organisations can build on the more nuanced guidelines that some have started to develop and test agents in more objective areas where they are likely to be strong — such as detecting errors.
They could also take steps to ensure that scientists view agents as “human-centric” tools that enhance, rather than bypass, their judgement — for example, by being transparent about the systems and ensuring that agents expose their reasoning, verify their claims with citations, and (to the extent possible) document their uncertainty. By understanding how an AI agent reaches its conclusion, scientists will have opportunities to learn, intervene, and collaborate.
These are formative years
The scientists using AI agents are not going back. The technology will continue to improve. The institutions that govern science — and the physical and social infrastructure it runs on — were built for a research enterprise that is now evolving faster than they are. We need to upgrade them for the agent era.
This will take a significant collective effort and serious investment. AI agents must not be viewed as a rationale to spend less on science; this would be a tragic false economy. The countries that treat this as a moment of genuine structural change will unlock the greatest public value and shape what comes next. The choices made in this window may be difficult to reverse.

[Оригинал](https://deepmind.google/public-policy/conjecture-machines-ai-agents-and-the-new-validation-bottleneck-in-science/)