When physicists recognised that an atomic bomb might be possible, they understood that a scientific discovery could alter the balance of power. They warned the President of the United States. The response eventually brought government, science and industry together through the Manhattan Project, on a scale far beyond normal academic research.
The institutional lesson is that a change in what a country can do may demand a change in how it organises itself. Waiting for existing institutions to absorb a new capability is not a strategy.
Agentic AI demands that we ask that question again. AI systems can already write and run code, operate software, use tools and carry out sequences of work. They remain unreliable, and the pace of further progress is uncertain. Yet the prospect of much cheaper and more abundant intelligent work is already concrete enough to require a response.
We should begin designing that response while the technology is developing.
Intelligence matters when it changes the world
Much of the debate about AI sovereignty concerns models: who trains them, who controls them and where the compute sits. These questions matter. A country's ability to use AI also depends on what it can discover, build and bring into use.
Can AI help us discover a drug, design a material or build a robot? Can it identify a decisive experiment, and can that experiment actually be run? Can a promising result be reproduced, engineered and manufactured?
Science is one of the principal ways intelligence becomes new capability. Countries and companies that learn to use AI to advance scientific frontiers could gain a compounding advantage across biology, chemistry, materials, robotics, agriculture and energy. The prize is both faster research and the technologies and industries it makes possible.
That requires laboratories, data, instruments, engineering, skilled people, funding and institutions that can coordinate them. Scientific sovereignty means retaining enough of this capacity to act when a new opportunity matters.
The UK has strengths across these areas, from its universities and national facilities to its life sciences sector and the NHS. The question is whether we can make them work together quickly enough to turn intelligence into discovery, and discovery into the capacity to act.
Science was built for a slower world
Scientific institutions evolved around the cost of analysis, writing, review and coordination. Finding the right expert, assessing a proposal or assembling a large collaboration took substantial human effort. Many of our processes still reflect those constraints.
Consider grants. A long application describes a project, but the effort of writing it also limits who applies. It makes a large set of possible projects manageable for a smaller set of reviewers. Whatever the formal purpose, labour helps ration access.
That mechanism becomes less useful when an agent can quickly produce pages of plausible prose. Automating the application may increase the volume without improving the decisions. We should instead ask what a funding system needs to achieve.
Imagine a standing account of the outcomes a field wants to reach, the capabilities it needs, the evidence behind those judgements and the bottlenecks blocking progress. Scientists could submit short proposals against those needs or challenge the account itself. They could identify something outside their own expertise that must be built: a dataset, assay, robot, software service, manufacturing process or new organisation.
Agents could find related proposals, prior attempts, contradictory evidence and missing skills. They could help assemble teams and turn an initial request into options that scientists and funders can inspect. People would remain responsible for priorities, scientific judgement, risk and the allocation of resources.
The change would reach beyond the grant form to the way a field chooses and organises its work.
Make better science affordable
Many improvements science needs are already understood. The Turing Way and the wider reproducibility community have developed practical guidance on version control, documented dependencies, reproducible environments, testing and shared code. A scientific analysis should be possible to inspect and rerun without access to its author's undocumented laptop.
Putting these principles into practice takes time. Repositories, tests, containers and documentation must be created and maintained. As software and computing platforms change, work that ran yesterday can fail tomorrow. Institutions often endorse reproducibility without providing enough support to sustain it.
Bioconductor and CRAN illustrate the value of shared scientific software and the continuing work it requires. A useful package may outlive the grant that funded it by many years. Maintaining it can become a lasting obligation for the scientist who wrote it.
Agents could reduce that burden by preparing dependency updates, diagnosing failed checks, reconstructing environments and updating tests and documentation. Maintainers would still need to check whether a proposed change preserves the scientific method and results. The opportunity is to make routine maintenance an affordable service that communities can support.
The same principle applies to review. CODECHECK provides a process for independently executing the computations behind a scientific paper and recording the outcome. This useful check requires skilled human time. Agents could assist with the first pass: fetching code, attempting to build the documented environment, running workflows and reporting missing files or failed steps. They should preserve a clear record of what ran and distinguish an author's original workflow from any proposed repair.
Peer review could become more layered. Domain expertise, statistical checks, software execution, reporting checks and searches for contradictory evidence address different kinds of failure. Agents could help carry out preliminary checks and route questions to the people qualified to judge them.
Communities, funders, journals and institutions should invest in freely available tools that make these practices easier to adopt. Reduced overhead alone will not solve failures of incentives or judgement. It could, however, make reproducible and well-maintained work a practical baseline for many more researchers.
Make science legible to agents
Scientific knowledge is recorded largely in prose. Claims, assumptions and relationships between pieces of evidence are often implicit. A reader must reconstruct the argument, and an agent can produce a fluent summary while misunderstanding what a result supports.
Science needs an additional computational layer alongside its narrative literature. Agents could help authors and reviewers make questions, claims, evidence, assumptions and disagreements explicit. Discourse graphs offer one emerging approach, connecting propositions to their supporting and opposing evidence. Such representations must preserve their sources, remain open to correction and allow competing interpretations.
Laboratory work needs a similar treatment. Paper records are hard to search, share or connect to computational tools. Electronic notebooks can help, but recording work is still an overhead when the researcher sees little immediate benefit.
Useful records should help scientists prepare experiments, detect missing steps, connect results to decisions and transfer methods between laboratories. Work on laboratory observability, including Transfyr, explores how to capture more of what actually happens at the bench. Instruments, video and wearable or augmented interfaces could help record and guide experimental work, including tacit knowledge that never reaches a written protocol.
These approaches also create risks of surveillance and misleading certainty. Scientists need consent, local control, ways to correct records and clear limits on their use. Better records should help people execute experiments and learn from them.
The aim is to make scientific knowledge easier for both people and agents to inspect, test and use.
Coordinate around problems that matter
Much of science is distributed across universities, companies, funders and countries without a shared management structure. A field may want the same broad outcome while lacking a common account of what blocks progress, who can help or where its important disagreements lie.
Keeping that account current is expensive. As evidence, capabilities and plans change, people lose track of dependencies and repeat work they did not know had already been attempted.
Convergent Research's Gap Map and Encode's AI for Science Challengescape, launched by Pillar VC, show how communities can make research obstacles visible. We can build on this approach by connecting problems to evidence, proposals, objections, people and capabilities.
Agents could help identify duplicated efforts, neglected problems and missing capabilities that block several desired outcomes. They could find people who should meet and keep track of what a team has tried. The organising principle would be the work a community wants to accomplish.
People must choose those outcomes. Agents could reduce the effort of coordinating around them and make disagreements visible early enough to change a plan.
If these tools are widely available, more laboratories could practise better science and contribute to shared goals. This is an opportunity to build useful scientific infrastructure across the research system.
Discover the institutions we need
We do not yet know what every institution of agentic science should look like. We need a practical effort to discover what works.
How should a field choose its priorities and identify the capabilities needed to reach them? When should a funder support a grant, procurement, shared platform, challenge, company, focused research organisation or national facility? What should a paper become when agents are readers and contributors? What should a laboratory make available to authorised agents and collaborators? Which capabilities should remain distributed, and which require national infrastructure?
We also need to reconsider credit. Scientific progress depends on maintaining, correcting and integrating work, as well as producing new findings. Institutions must recognise those contributions if they want people to make them.
Early answers are being developed in separate communities. The task is to connect them, test them and help successful approaches operate at a useful scale.
That requires social as well as technical change. People need enough agreement to act while keeping uncertainty and dissent visible. A field should be able to state its competing views, expose the assumptions behind them and identify the evidence or experiment that would change the argument. Prestige should not determine which position becomes the official account.
This work crosses science, government, philanthropy and industry. It needs sustained coordination between them.
What the Network is for
The UK Agentic Science Network should provide a community-led, independent forum for this work. It should connect existing communities and create routes to participation and leadership across the whole UK, beyond the usual institutions and invitation lists.
Its purpose is to help people design and test how science should work in the agentic era. We propose four tasks.
Identify where change is required
Find the practices, institutions and missing capabilities that could limit progress as agents improve. Grants, publishing, laboratory automation and conferences are starting points. Other priorities will emerge when people from different disciplines compare their experience. Each should be grounded in a specific obstacle to useful scientific work.
Publish calls to arms
For each selected area, publish a short argument explaining what is changing, why the current arrangement falls short, what an alternative could look like and which questions remain unresolved. Give people a proposal clear enough to criticise, correct or build upon.
Build a public account of what needs to be done
Bring scientists, engineers, founders, funders and operators together through events around the country. Use agents to help compare contributions, identify recurring claims, preserve minority positions and find people whose experience is missing.
Publish the arguments, disagreements and revisions wherever confidentiality allows. Participants must be able to correct how their views are represented. The result should be a developing account that people can inspect and use as they decide what to do.
Turn the argument into work
Help form teams, connect them to the institutions they need and bring relevant funders and partners into the discussion. The right response may be a collaboration, grant, procurement, standard, shared service, company, focused research organisation or national facility. Its form should follow the problem.
Each major strand should move towards a named owner, a willing team, a concrete next step and a way to judge whether it worked. A gap with no owner is a complaint. The Network should help turn gaps into plans and plans into capabilities.
Experiment on ourselves
The Network's own operation should be an experiment in agentic science. Over the coming year, we propose a programme of arguments, local meetings and practical work leading to an agent-first national conference.
Before the conference, participants would contribute positions, evidence, capabilities and unresolved questions. Agents could help synthesise this material, identify conflicts and prepare questions that deserve time together.
At the conference, that time should be spent on structured debates, working tables, design sessions, demonstrations and team formation. Presentations that people can watch beforehand should be available beforehand. Where a question cannot be resolved in the room, participants should identify an experiment, analysis or build that could resolve it.
Afterwards, agents could help maintain a public, versioned account of what changed, track plans and dependencies, and support follow-through. People would remain responsible for every attributed position, commitment, funding decision and public claim.
Success means people leave with a clearer understanding of the problem, teams ready to work and specific next steps they have agreed to take.
Our proposed commitments
- Take agentic AI seriously. Prepare scientific institutions while the technology develops, without waiting for certainty about its final form.
- Focus on real-world capability. Judge progress by what science can discover, build and learn.
- Redesign intentionally. Start from the outcome required and reconsider the process used to reach it.
- Reason in public. Publish arguments, uncertainties, disagreements and revisions wherever confidentiality allows.
- Make disagreement productive. Preserve minority positions and seek evidence or experiments that could move the argument forward.
- Organise nationally. Make participation possible across the UK, including beyond established networks.
- Coordinate for execution. Give each major strand an owner, a team, a next step and a test of whether it worked.
- Experiment on ourselves. Evaluate and improve the Network's events, tools, use of agents and governance.
- Keep people accountable. People retain authority over scientific judgement, representation, commitments and resources.
A call to arms
Without deliberate change, AI could simply accelerate the production of papers, proposals and disconnected activity. The UK could gain access to powerful models while remaining dependent on others for the laboratories, infrastructure and institutions that turn them into new capabilities and industries.
We can choose a more ambitious response: decide which outcomes matter, identify what must be built and organise people and resources around that work.
We need scientists who can explain where the present system fails, engineers who can build what is missing, founders prepared to create new organisations, funders willing to rethink how work is selected and institutional leaders willing to test changes in practice.
Bring a bottleneck you understand, an argument you think we have wrong or a capability you can help build. Help us define a tractable first step, the people it needs and the evidence that would show progress.
Our ambition is to make the UK a place where AI advances frontier science and where those discoveries become capabilities and industries we can build and sustain.