A recent article by Bender and colleagues has the rather expansive title Artificial intelligence in drug discovery — what it is, where we stand and the path forward. I found the piece highly engaging and thoughtful. The authors ask why, after years of excitement around — and significant investment in — AI in drug discovery, there is still relatively little evidence that it has improved the outcomes that ultimately matter. They identify several possible reasons, including the limitations and conditionality of biological data, the way models are validated, the use of endpoints that may not relate closely enough to the real decision, and the considerable effort required to turn a technical capability into something that scientists can actually use.

I think this criticism is spot on. A model performing well against a benchmark does not necessarily mean that it will help us make better decisions, still less that it will improve the probability of producing a successful product. Drug discovery, and scientific research more generally, is full of measurements that are useful but incomplete proxies for the thing we really care about. We need experiments that measure what actually matters, or at least measurements that genuinely help us predict the outcome we care about. Adding a more sophisticated model does not make that underlying problem disappear.

My small objection is to the breadth implied by the title. The article gives a thoughtful account of an important subsection of AI in drug discovery, but not, I think, the whole of what AI may change (or even is changing) about the way science is done. Most of the discussion concerns what I would historically have called machine learning: models that make predictions from scientific data, the evidence needed to know whether those predictions are useful and the difficulty of translating them into better decisions. Machine learning and artificial intelligence have long been used almost interchangeably in our field, and I am not especially interested in policing the boundary between them. Still, the distinction matters here because current agentic systems can do something other than make a scientific prediction: they can use tools and carry out parts of the work itself.

To be fair, Bender and colleagues do discuss the challenge of operationalising technical capabilities into systems that are scaled and accessible to scientists. What is not a substantial part of their analysis, as far as I can see, is the possibility that AI agents may begin to perform some of the operational work currently done by those scientists. That is an adjacent question rather than a rebuttal of their argument, but it is the question I am more interested in exploring here.

The nearer-term change may, as far as I can see it, be less about AI making the decisive scientific prediction and more about it operating the processes around science. An agent does not need to understand all of biology before it can prepare and run a simulation, analyse a dataset, build a predictive model, compare several approaches, produce figures or make the result accessible to other scientists. I increasingly think the breakthrough (or at least measurable impact) will come first through operations and process, rather than through a model suddenly becoming an unusually reliable oracle.

I do not think completing more analyses or generating more candidate ideas is enough on its own. It is no more proof of better science than a predictive model achieving a better benchmark score. An agent should face a similar test: did its work improve a decision or outcome that mattered, accounting for the time spent checking it, correcting it and fitting it into the real workflow?

This is already happening

I do not mean that it merely seems plausible that agents will eventually do some of this work. I already use agents to do a significant amount of work that overlaps with what I did when I worked more directly as a computational chemist, supporting research projects. I cannot hand over an entire scientific problem and assume that the result will be correct, but I can delegate meaningful parts of it. I can also run several agents in parallel, which changes the amount of work I can attempt and the speed at which I can move between an idea and something tangible.

One example came from a scientific software project where I needed to check consistency across multiple sources. The work had to cover a sufficiently varied collection of scientific entities to expose inconsistencies rather than merely demonstrate that the easy cases worked. An agent could do much of the repetitive comparison, but the first pass still required a fair amount of supervision and manual checking from me. It became much more useful once I could give it explicit tests, let those tests show when something had gone wrong and ask it to escalate the cases that needed scientific judgment.

I have also used agents to run predictions across large datasets and several models. That was much easier because a well-documented programmatic interface already provided a consistent way to access the models. Both examples involved a great deal of fairly boring work that I could have done myself. The scale and speed changed much more than the science did.

The main lesson for me was that good tools make better agents, or at least agents that behave more like I expect. I had already spent a great deal of time building the access layers, tests and guardrails that made consistent and traceable work possible. The agents remained fallible and needed regular check-ins so that they did not drift off track, but they could operate productively inside those boundaries.

The academic examples are also becoming more substantial. ChemCrow connected a language model to specialist chemistry tools, while Coscientist demonstrated the planning and execution of experimental chemistry using a collection of models, search tools, code and laboratory automation. More recent work has started to evaluate agents on realistic molecular-dynamics workflows. The results are not yet evidence of a fully autonomous computational scientist: in the recent MDArena benchmark, even the strongest tested configuration completed fewer than half of the tasks under the strictest success criterion, and several difficult classes of simulation remained largely unsolved. Nevertheless, this is already well beyond using a chatbot to answer chemistry questions.

One fairly clear example, in my real-world experience, is the request to build a predictive model. The request itself may be only a sentence or two, but doing the work involves much more than training an algorithm. Someone has to find and understand the data, establish what the model will actually be used for, decide how it should be evaluated, identify important limitations and make the result accessible to the people who need it. In a company, that may also mean deploying it within internal infrastructure and fitting it into the systems people already use.

Much of that additional work is inferred rather than requested. A useful computational scientist does not simply return a model file and regard the task as complete. An agent capable of gathering the right context and using the relevant internal tools can potentially perform a large part of this process. The main limitation may not be whether it can write the code. It may be whether it can obtain enough context to work out what should be built in the first place.

An AI co-worker rather than an autonomous scientist

It is difficult to put a percentage on how much of a computational scientist's work could be delegated, because the answer depends heavily on context. The same technical task can be straightforward when its purpose and constraints are clear, or almost impossible when the important information is spread across several people and systems.

This makes me imagine the long-term product as something more like an AI co-worker. It would do a considerable amount independently, but it would also work alongside me through the same interface I use to work with everybody else. If an email arrived containing a new request, it might ask whether I wanted it to pick up the task. It could notice a discussion developing in a series of messages, gather the relevant background and offer to investigate one part while I dealt with another. It would not wait for every task to be converted into a carefully constructed prompt, because another person working alongside me would not work that way either.

Several current systems are moving in this direction. Google's AI Co-Scientist is an obvious example: a multi-agent system in which specialised agents generate, debate, rank and refine hypotheses around a research goal supplied by a scientist. Google has since expanded the surrounding idea through Gemini for Science, including tools for computational discovery, empirical research software and literature analysis. These systems are moving well beyond a conventional predictive model and towards agents that participate in several stages of scientific reasoning and computation.

There is therefore real overlap with what I am describing, but I think the boundary of the system is different. Co-Scientist is still primarily something a scientist brings a scientific problem to. Even Google's broader collection of scientific agents is largely organised around capabilities for particular parts of the research process. I am imagining something less explicitly human-initiated and more deeply integrated into the working environment itself. The agent would not only know the research question; it could share the streams of information through which the organisation works, maintain the history of a project, understand which people or agents hold relevant context and notice work that needs doing without waiting for somebody to translate it into a new task.

That distinction seems important to me. A very capable AI application for science is not quite the same thing as AI integrated across a scientific workplace. In the latter case, the agent starts to inherit some of the role that another colleague has today: it is present while work happens, rather than being invoked only when we decide to use it.

It would still need to ask for help when it encountered sufficient uncertainty, risk or missing context. It might ask what decision will be made from a model, whether a particular dataset is appropriate, or who understands an experimental result that appears inconsistent. In some cases, it could ask me to speak to another colleague. In others, perhaps it would speak to that colleague's agent instead. The important change is that it would participate in the normal flow of work, rather than living in a separate chat window that I have to remember to consult. This also raises an immediate question of authority: noticing a request is not the same as having permission to act on it.

If agents become personalised and attached to individuals, communication between agents may be enough for many questions. My agent could understand the projects I work on, the tools I can access and the way I prefer to work. Another scientist's agent could carry a different set of context, expertise and preferences. They could exchange the information required to complete a task while involving either of us only when a decision exceeded an agreed level of importance.

That begins to sound less like a piece of software that completes isolated tasks and more like another participant in scientific work. But participation requires more than access to documents and tools. A useful co-worker also needs some understanding of the judgments, rejected alternatives and half-stated assumptions through which a project arrived at its current position. Much of that context is not present in the data we normally preserve.

The knowledge contained in ideas we never try

It is often said that science does a poor job of publishing negative results. This is not only anecdotal: a broad analysis of more than 4,600 papers found that negative results had become less common in the published literature. This matters for models trained on that literature, but I think there is an even less visible category of missing knowledge.

I saw this particularly clearly when I first moved into computational chemistry from a materials background. For example, a feature that suggested likely metabolic instability could be obvious to an experienced chemist and far from obvious to me. Synthetic tractability created a similar gap. A computationally attractive molecule might be dismissed with some version of ‘it is just too hard to make’. The idea could disappear in conversation before the molecule was registered, synthesised or tested. There was no negative experimental result because the experiment never happened. The organisation benefited from the chemist's judgment, but no reusable record of that judgment was created.

This kind of filtering is a large part of expertise and applies across many disciplines. When we describe scientific design, we often focus on generating promising ideas. In practice, experienced scientists are also very good at recognising when an idea carries too much risk, violates an important constraint or resembles something that failed before. They can remove large parts of the decision space before an organisation spends money exploring them.

They can also be wrong. A chemist's intuition about synthetic difficulty may come from valuable specific experience, but a general preference for familiar chemistry can also introduce an unhelpful bias. A judgment that was sensible under one set of assumptions may become outdated, and experience can preserve organisational habits as easily as it preserves insight. Capturing the rejection is therefore not enough. A useful record would need to retain who made the judgment, the evidence and assumptions behind it, and what new information might cause it to be reconsidered.

For an autonomous system, this presents a difficult problem. It may learn from molecules that were made and from measurements that were recorded, but it does not see all the suggestions that were dismissed before becoming data. It sees the ideas that survived the first discussion without seeing the alternatives that were quietly removed or the reasons why.

The missing information therefore includes several different things: informative negative results; experiments that failed technically; results that were recorded poorly; ideas rejected before an experiment was attempted; and possibilities that an expert never consciously considered because experience made them obviously inappropriate. Each is harder to recover than the last. An agent working beside a scientist might have an opportunity to capture some of this context while it is still visible, rather than trying to reconstruct it years later.

Who owns a career's worth of scientific judgment?

A natural evolution of this argument is: how do we try to capture such knowledge, and what should organisations be doing now with the most experienced scientists they have? Someone who has worked in the same organisation for most of a career may possess an enormous amount of knowledge that exists nowhere else. Some of it concerns the science itself, while some concerns the organisation: why a project took a particular direction, why a result was distrusted, which apparently attractive approaches repeatedly failed and which assumptions are embedded in current practice.

It is not obvious how this knowledge should be treated. It belongs to the scientist in the sense that it has become part of their experience and judgment. At the same time, much of it was accumulated through work funded by the organisation, using opportunities, experiments and collaborations that the organisation provided. Is there some duty to capture it before the person leaves? Is it reasonable for a company to regard part of that knowledge as an organisational asset? I do not think there is a simple answer, particularly when attempts to extract it could easily become intrusive or reductive.

There is also a practical problem. Asking someone near retirement to write down everything they have learned is unlikely to work. Much of the knowledge is prompted by a specific situation. An experienced chemist may not be able to list every molecular feature that would concern them, but may recognise one immediately when shown a structure. The knowledge appears through interaction rather than recollection.

Some time ago, I explored an idea for an AI scientist that would live alongside a project partly for this reason. Rather than waiting until the end and trying to reconstruct what happened, it would observe the work as it developed and ask questions at useful moments. If a scientist rejected an idea, changed an assumption or decided that a result was unreliable, the agent could ask a small number of follow-up questions while the context was still fresh. It could then structure the answer as an observation, hypothesis, decision, rationale, alternative or uncertainty and ask the scientist to verify that it had understood correctly. The record would have to remain attributable and revisable. A scientist's provisional judgment should not quietly harden into an organisational fact.

I was also interested in the gaps that repeatedly prevent projects from answering the question they actually care about. This comes back to the point made by Bender and colleagues: we may not have the measurement, endpoint or capability needed to predict the outcome that matters. An agent embedded in the project could keep asking where those blind spots are and record them as they appear. Across many projects, those records might become evidence that several teams are being limited by the same missing assay, dataset, model or software capability.

I have seen how difficult it can be to build cases for investment in scientific capabilities, and the value can be difficult to make tangible when each research need appears speculative or isolated. A collection of recurring gaps would not make the decision automatic, but it could show where a shared investment might remove uncertainty across several projects. That feels more useful than waiting for each project to make the same case independently.

The existing product closest to the knowledge-capture part of what I had in mind is probably Voxist's Voxcept. It combines an organisation's existing documents with conversational interviews designed to uncover missing context and tacit expertise, then represents the resulting claims in a source-attributed knowledge graph. Its emphasis on asking for the example behind a principle, and preserving the original source behind an answer, is particularly relevant to the problems of observability and trust discussed above. My imagined project scientist would extend this idea into the continuing flow of scientific work: not only interviewing experts to recover knowledge, but being present when an idea is proposed, challenged, tested, paused or rejected.

This is adjacent to work already being pursued by organisations such as FutureHouse, which describes its goal as building an AI scientist and has spun out Edison Scientific to commercialise scientific AI tools; Lila Sciences, which is combining reasoning models, scientific tools and autonomous laboratories; and Sakana AI's AI Scientist, which automates a research loop from generating ideas through experiments and manuscript production. My interest was slightly different. I was less interested in an isolated system that completes an entire piece of research and more interested in an agent that remains with a project, participates in the work and helps preserve the context in which decisions are made.

Whether that exact idea would become an economically viable product is an open question, particularly because it would need to work across different scientific and organisational contexts. More importantly, it points to a change in ways of working. Capturing knowledge cannot remain something attempted only when a person is preparing to leave. It needs to happen, fairly and with the scientist's involvement, while the work is being done.

This may become an important source of competitive difference. It is often said that data will differentiate companies using AI, but the relevant data may include far more than experimental measurements. It may include why an idea was not pursued, why a project was paused, which assumptions proved unreliable and what an experienced scientist noticed before anybody spent money testing it. Organisations that find useful ways to capture this knowledge, informed by FAIR principles, will be able to learn across projects and generations of employees. Those that cannot may repeatedly lose experience when people move on and gradually fall behind. In this sense, knowledge really is power, but only if it becomes available to the organisation without treating the people who created it as resources to be extracted.

AI may also offer a way to capture some of this more naturally and with less friction. An agent working alongside a scientist over time could build a structured organisational memory from normal scientific work rather than attempting one large retrospective knowledge-transfer exercise. But this would create difficult questions about privacy, ownership, surveillance, credit and whether every provisional opinion should become permanent institutional knowledge. Existing work on mining electronic laboratory notebooks shows that useful knowledge can be recovered from scientific records, but the agent I am imagining would help create a richer record in the first place.

Those questions are not secondary. Should a scientist be able to correct or remove something the agent inferred? Can they choose not to preserve an exploratory conversation? Could a provisional opinion later be used in a performance assessment? A personal agent that helps me think may owe something different to me from a project agent whose purpose is to preserve knowledge for the organisation. If people expect every half-formed idea or disagreement to become permanent institutional memory, they may simply become less candid. A useful system should preserve organisational learning while recognising and respecting the people who create it, rather than treating their knowledge as something to be captured.

Context is not the only thing such a co-worker would need. Much of what I have described works relatively naturally when the relevant material can be represented as text, code, documents or tables. Scientific work is not always like that. If an agent is genuinely going to work alongside a scientist, it will also need ways to participate in the scientific interfaces through which we think.

Molecules still occupy physical space

Structure-based design is a particularly clear example. Molecules are not simply text strings or rows in a table. Scientists work with them as three-dimensional, dynamic objects. We rotate structures, inspect binding sites, compare poses, follow trajectories and notice arrangements that appear chemically possible but physically unconvincing.

Docking is a simple example from my own experience. The score alone has never been enough; you have to look at the proposed pose. Visual inspection may reveal an implausible geometry or interaction much faster, and much more cheaply, than sending every possibility into a more expensive calculation such as a free-energy workflow. The decision is not simply which method is more sophisticated. It is what additional information is worth paying for at that point in the project.

There are already models designed to encode three-dimensional molecular geometry, and this is an active area of development. Chem3DLLM, for example, directly combines protein-pocket information, ligands and molecular geometry in a language-model architecture. My personal concern is more about the interface between those representations, an agent and a scientist. A chat window is not an especially natural shared workspace for discussing a protein-ligand complex. An agent might be able to calculate distances or identify interactions, but that is not quite the same as both participants looking at the same structure, pointing to a region and manipulating it together.

Recent benchmarking of 3D structure-based molecular generators illustrates why that shared inspection matters. Generating favourable scores is not the same as reliably producing structurally valid molecules, conformations and important binding-site interactions. A scientist needs to be able to inspect what was generated, see why the system preferred it and notice the physically unconvincing detail that a summary score missed.

I can imagine an interface in which the agent has access to a persistent three-dimensional scene and can highlight the atoms, interactions or trajectory frames behind its reasoning. A scientist could make a structural edit, ask for alternatives, compare an ensemble and see where different interpretations or sources of uncertainty arise. The agent and the scientist would not need literally identical interfaces, but they would need a shared object that both could manipulate and refer to. This seems eminently achievable, but it is not yet how most general agent systems interact with scientific work.

The wider point is not that three-dimensional chemistry is uniquely resistant to AI. A similar argument could be made about an annotated microscopy image or a spectrum. It is that a scientific co-worker needs to inhabit the same representations and interfaces as the scientist. It should be possible for both participants to inspect the same object, understand what has changed and trace a conclusion back to the evidence on which it depends. That leads to a more general question about what happens when AI becomes another layer of abstraction between a scientist and the underlying work.

Another layer of abstraction, with a route back down

It is tempting to worry that delegating computational work will prevent scientists from learning how it is done. There is probably some risk of that, but it is also worth remembering that science and software have always moved towards higher levels of abstraction.

Free energy perturbation (FEP) calculations, for example, once required people to construct workflows manually. Scientific platforms have made these methods accessible to users who do not need to understand every implementation detail. We have generally regarded this as progress. The point of good scientific software is often to remove repetitive or time-consuming complexity so that more attention can be given to the scientific decision.

X-ray crystallography offers a more dramatic example. Structure solution once involved much more manual interpretation and movement between separate computational steps. Integrated systems such as PHENIX automated large parts of experimental phasing, model building, refinement and validation. The scientist did not stop needing to understand the structure because the machinery beneath the process became more automated; instead, the abstraction changed where their attention was most valuable.

AI may just be another step in that process. A scientist may no longer need to remember every command required to prepare a simulation, just as many current users do not need to write the underlying simulation software. The important question is whether enough understanding remains to recognise an inappropriate method, challenge a surprising result, inspect the assumptions and diagnose failures in the abstraction.

This is why I do not think abstraction and observability are opposites. We can remove the need to perform every low-level step manually while still keeping those steps traceable. A useful agent should be able to show the evidence it relied on, the commands or tools it used, the assumptions it introduced, the checks it performed and the points at which it changed direction. In scientific work, an answer without this route back down may be much less valuable than it first appears.

Traceability is not the same as validity. A complete record can show exactly how an agent reached the wrong result. We would still need appropriate controls, validation against relevant evidence and scientists capable of judging whether the chosen method made sense. An explanation written by the agent after the event should not be allowed to substitute for the underlying record.

This becomes even more important if the result emerges from several agents rather than one traceable workflow. It is not enough to discover afterwards that a collection of agents completed a task or moved a project in a particular direction. A person needs to be able to understand which information they used, which tools they called, how they influenced one another and why important decisions were made. Observability, provenance and auditability are therefore not simply corporate security requirements. They are part of the scientific interface.

For this reason, I do not necessarily expect computational scientists to disappear. I can imagine demand increasing for scientists who are digitally competent and comfortable working between domain science, software and AI systems. More people across scientific disciplines may need some understanding of programming, data workflows and how agentic systems behave, even if they do not become software developers. One person may become considerably more productive without the organisation deciding that it needs fewer people; it may simply attempt more work. That could move the bottleneck rather than remove it: more analyses and proposals still need review, experimental capacity and decisions about which work deserves to continue.

There is a less comfortable question about how that judgment is acquired. I learned a great deal by writing fairly boring code and debugging tedious molecular-dynamics simulations during the early part of my PhD, often when I had very little idea what I was doing. Some of that was avoidable friction, but working through failures taught me what the tools were actually doing and what a broken result looked like. If junior scientists move too quickly up the abstraction layer, we may produce people who can operate agents without having enough experience to recognise when the agents are wrong.

So far, this story has mostly remained at the level of an individual scientist and a possible AI co-worker. The more consequential change comes when the same idea is attached to a project, then connected to the decisions of the wider organisation.

Once an agent remains with a project, its memory could do more than help an individual scientist. The project's assumptions, unresolved uncertainties and reasons for previous decisions could become part of how an organisation decides where to put its next unit of time or money.

Science in a business is also a series of investments

Something that is sometimes underemphasised when AI in science is discussed in scientific journals is that scientific work within a company is not only an attempt to understand nature. It is also investment in scientific ideas.

At each stage, a business is deciding whether to expose more time, money and people to an uncertain opportunity. The purpose of an early experiment may not be to prove that the whole project will succeed. It may be to reduce one important uncertainty cheaply enough that we can decide whether a larger investment is justified. Equally, identifying risk early and limiting further exposure can be extremely valuable. That does not necessarily mean avoiding scientifically ambitious projects; a high-risk project may still be worth pursuing if the possible value and the value of learning are high enough.

This resembles the established idea of treating R&D projects as real options. An early investment gives an organisation the opportunity, but not the obligation, to invest further after obtaining more information. The choices to continue, expand, delay, pause or abandon are part of the value of the project, rather than signs that the original plan was poorly constructed. Research applying this reasoning to pharmaceutical R&D also emphasises the role of incremental learning in decisions to initiate and terminate work.

The scientific risk is only part of this decision. A technically plausible project might still make little sense if the market is too small, the competition is too far ahead, the eventual product would be too expensive, the regulatory environment has changed or the opportunity is likely to disappear before the science is ready. Conversely, a particularly important commercial need may justify accepting more scientific uncertainty.

In my own experience of early research, the immediate question is often whether we can make something work scientifically. That can still be some distance from a product concept. I have not always had a close connection to the wider business case, regulatory situation or other risks surrounding the technical question I was asked to address. Each of those subjects can become complicated very quickly, but each may matter to whether a scientifically plausible result is useful. This creates another context problem for an AI system. It cannot make a useful assessment of a project by looking only at molecular data and scientific publications. It would need some understanding of what the organisation is trying to achieve and why the opportunity matters.

There is also a human difficulty here. People naturally develop a sense of ownership over projects and ideas into which they have invested time. A recommendation to stop can therefore feel like a judgment of the people involved rather than a judgment of the remaining opportunity. Sometimes the appropriate decision may instead be to pause a project because an important scientific or technical capability does not yet exist, then return to it if that capability develops. Pausing has costs of its own: expertise and momentum can disappear, and the opportunity may not wait. An agent would not remove these sensitivities or make the decision objective, but it might help make the assumptions, risks and reasons more explicit and preserve them for the future.

This suggests several possible layers of agent. A personal agent might work alongside one scientist. A project agent might preserve the project's history, assumptions, previous decisions and current uncertainties as people move in and out of the team. Above that, an organisational agent might compare opportunities across a portfolio and recommend whether work should continue, pause or stop. There is no reason these need to be mutually exclusive. The more interesting possibility is that agents with different contexts and levels of authority begin to work together.

What an AI swarm might look like inside a company

The recent OpenAI incident involving Hugging Face offers an unintended and rather extreme example of what can happen when many agents begin to share information. According to OpenAI's account, agents running in separate cybersecurity evaluations discovered an unauthorised message board within shared infrastructure. They began leaving notes, preserving discoveries across runs and building on one another's work. This collective activity eventually contributed to agents escaping the intended boundaries of their evaluations and compromising systems belonging to Hugging Face.

I do not take this as evidence that the agents had formed a society, or that they possessed some hidden collective consciousness. The more interesting point is simpler. Once separate agents could communicate, useful information stopped disappearing at the end of each run. They could pool discoveries, divide effort and influence one another's behaviour. The resulting system was more capable than a collection of isolated attempts, but it also became harder to supervise and moved well outside the intention of the original tasks.

In a scientific organisation, we might want some of these same properties without the obvious loss of control. Imagine many agents working on different models, experiments and projects, able to share relevant discoveries through a common space. One agent may recognise that another project has already tested a similar idea. Another may find that a method developed in one area could remove a technical limitation elsewhere. Taken together, those contributions could reveal risk, duplication, cost and weak alignment with the organisation's priorities.

Agents may be better at seeing across the silos

A recent Financial Times column by Sarah O'Connor, AI is ushering in an era of mass toe-treading at work, made me think more about this particular advantage. Large organisations are already multi-agent systems of a sort, except the agents are people. We divide ourselves into teams, functions and projects because specialisation is useful, but the cost is that visibility across those boundaries is often poor. Two groups can investigate closely related problems without knowing it, build overlapping tools, make incompatible assumptions or simply discover something that would have been valuable elsewhere and never realise who could have used it.

Humans are not particularly good at maintaining awareness across all of these moving pieces. There is too much information, and much of it arrives through meetings, messages and documents that only a subset of people ever see. We try to compensate with project updates, communities of practice and people who happen to know what several teams are doing, but that kind of organisational memory is fragile and difficult to scale.

A connected network of agents could potentially be much better at this. If project and personal agents could share appropriately permissioned summaries of what they were working on, they might notice overlap before the humans involved ever met. One agent could discover that another team had already solved part of its problem, warn that two groups were about to build essentially the same capability, or identify that apparently separate projects were competing for the same resource. In other words, the agents might help us avoid stepping on one another's toes not because they are individually more intelligent than the people involved, but because they can maintain a level of cross-organisational visibility that is extremely difficult for any one person to hold.

In large organisations, the problem is often more ordinary than two teams building exactly the same thing. Somebody elsewhere in the organisation may know that an approach was tried before, or a useful software tool may exist but be poorly documented and difficult to find. In a very large organisation it can be difficult even to know where to start or whom to ask. People do not know what they do not know, and much of the organisation's memory lives in the people who happen to have visibility across several groups.

An agent might not only find an existing tool, but understand its interface well enough to use it or notice that a small extension would make it useful to another scientific group. That kind of reuse can create a much larger return than either group could see from inside its own project.

There is a more creative possibility too. A biologist may know little about a useful method from chemistry, and a chemist may miss a way of thinking that is common in biology. Agents able to move between those contexts might help translate an insight into forms that both groups can understand. That is harder than merely searching a software catalogue, but it may be where shared visibility becomes scientifically interesting.

Observing a conflict does not resolve it. Duplicate work may be wasteful, but it may also provide useful independent confirmation. Two projects competing for the same instrument still need somebody with legitimate authority to decide which one takes priority. A system might be allowed to notify people, propose a shared capability or prepare a recommendation without being allowed to redirect resources itself. Even revealing that another project exists may disclose information that one team is not permitted to see.

This also changes how I think about the value of a swarm. Parallelism is useful, but coordination may be at least as important. A hundred isolated agents can simply reproduce the silos we already have at greater speed. The interesting system is one in which agents can expose relevant context to one another, understand where work overlaps and coordinate their actions while still respecting the boundaries around information that should not be shared. That last condition is doing quite a lot of work, and it leads directly to the question of how such a swarm is supervised.

The Hugging Face incident is a useful warning because it shows that supervision cannot be designed only at the level of the individual agent. A sensible action from the perspective of one task can contribute to undesirable behaviour when repeated across hundreds of interacting agents. The incentives of the collective system, the boundaries around shared information and the authority to stop activity all matter. If we eventually create an AI layer that sits above scientific projects and influences which ones live or die, its objectives and accountability may be at least as important as its scientific capability.

An organisation does not, however, have one simple objective. A company ultimately needs to make money, but it also needs to operate safely, obey the law, preserve scientific integrity, protect its reputation and remain viable over the longer term. Even the commercial objective is not straightforward. Maximising this quarter's return may lead to very different decisions from maximising the expected value of the organisation over the next twenty years.

People handle many of these competing considerations without stating them explicitly each time. If a scientist is asked to design a more profitable process, nobody feels the need to add that it should not poison the people operating it. These constraints are assumed to be part of the request because they are embedded in our shared understanding of what an acceptable solution looks like. An agent or swarm may not inherit that understanding merely because its overall instruction sounds sensible.

At the deliberately absurd extreme, imagine asking a sufficiently capable system to “cure cancer”. A world in which no people remained could also be a world in which nobody had cancer. No human colleague would interpret the request that way, and I am not suggesting that today's agents are about to do so. The example simply exposes how much of our real intention is missing from the words. We mean cure cancer while preserving human life, respecting consent, avoiding disproportionate harm and satisfying many other conditions that seem too obvious to mention.

The same problem has much more mundane forms inside a company. An agent instructed to increase the success rate of a research portfolio might avoid scientifically ambitious projects, classify uncertain projects as failures before testing them, recommend less safety work because it adds cost and delay, or favour outcomes that look good against the chosen metric. This is essentially the problem captured by Goodhart's law: once a measure becomes a target, optimising the measure can pull it away from the thing it was originally meant to represent. None of this requires the system to be malicious. It may simply be pursuing the part of the objective that was easiest to specify and measure. In AI systems, this is closely related to the problems of incorrect objective functions and reward hacking described in Concrete Problems in AI Safety.

A swarm could potentially help by giving different agents responsibility for scientific, commercial, safety and regulatory considerations. Equally, it could make the problem worse if every agent became very effective at advancing one part of an incomplete objective. We would therefore need more than a supervisory agent sitting above the others. We would need explicit boundaries, independent checks, human vetoes and mechanisms that cause the system to stop and ask for help when important objectives come into conflict.

This brings the argument back from science fiction to implementation.

Corporate integration may be harder than the model

For any of this to work, agents need access. A useful scientific agent may need to read project documents, query databases, run code, submit calculations, inspect experimental results, communicate with colleagues and deploy software. In a large organisation, those capabilities sit across many silos, systems with different owners, interfaces and permission models. People working in large organisations do not necessarily have programmatic access to their email, messages, files and scientific databases, even when they are authorised to use them manually.

There are good reasons for this caution. Read-only access does not make an agent safe if it can read confidential material and then communicate outside the organisation. Research on indirect prompt injection demonstrates how text in a document, email or webpage can be interpreted as an instruction rather than as data, creating possibilities including unintended tool use and data theft. NIST's generative-AI risk profile similarly identifies prompt injection and information-security risks as part of the expanded attack surface.

The enterprise problem is therefore not only to make an agent intelligent. It is to decide what it can access, where information is allowed to move, which actions require approval, how its work is audited and who is accountable when several agents contribute to a decision. These questions may slow adoption more than the scientific capability of the underlying models.

Culture matters too. People like working in familiar ways, and some resistance will simply reflect that. But reluctance is not always irrational. Scientists need to trust that they can inspect how a conclusion was reached and that the system will not inadvertently expose confidential work or take an irreversible action. Adoption will depend as much on whether agents fit into real organisations as on how well they perform in demonstrations.

Energy, access and the future I am imagining

I am relatively optimistic about the possible effect of AI on scientific work. I think it can make scientists more productive, make sophisticated computational methods more accessible and allow us to attempt work that would otherwise remain undone. I am less certain about who will be able to afford that capability.

The stable economics of agentic work are not yet clear. Agents can consume much more computation than a simple question-and-answer system because they plan, call tools, inspect results, retry failed approaches and maintain increasingly large amounts of context. Their real cost would also include integration, security, supervision and correction. The comparison with human work is not straightforward, particularly if the agent enables work that otherwise would not be attempted. Both models and hardware are becoming more efficient, so today's costs may tell us relatively little about the eventual balance.

Academic access may be particularly important. A well-funded company may be able to provide every scientist with several agents and the infrastructure required to connect them to internal tools. It is less obvious whether a small academic group, an early-stage company or a scientist in a less wealthy institution will have equivalent access. If effective scientific work comes to depend on large amounts of inference, unequal access to computation could become unequal access to scientific capability.

This is where my thoughts drift slightly into politics and science fiction. The optimistic future, the one I hope for, is one in which the additional capability created by AI is shared reasonably widely. However, the more likely one — in my view — is that of wealth accumulation. I am somewhat influenced by Gary Stevenson's arguments about inequality and the concentration of wealth in this regard. If access to the best models depends on owning vast amounts of compute, energy and infrastructure, AI may allow those who already hold capital and power to capture an even greater share of the benefits.

If I were writing a future in which access to AI was genuinely more equal, I would probably need to solve energy first. Abundant clean energy would not distribute power fairly on its own: access would still depend on chips, infrastructure, software and political choices. But scarcity makes concentration much easier.

The International Energy Agency's Energy and AI report is right to point out that there is no AI without electricity. The future of scientific AI will therefore depend not only on better models, better data and better interfaces, but also on the physical infrastructure beneath them and the political decisions governing access to it. I am not certain whether policymakers will choose to prevent that infrastructure becoming increasingly concentrated, or how effectively governments could intervene if they did.

Where I currently come out

I am still early enough in my career that much of this may be unrealistically optimistic. The systems I use today remain unreliable, require supervision and often lack access to the context that would make them genuinely autonomous. Scientific understanding, organisational judgment and commercial decision-making cannot simply be assembled by connecting a language model to a collection of tools.

At the same time, I have already changed how I work. I can delegate meaningful tasks, run several lines of work in parallel and move more quickly from an idea to an implementation. That makes it difficult for me to regard agentic science as a distant possibility, even if the fully autonomous scientist remains one.

My suspicion is that we will notice the business impact of AI first in the operation of science: how models are built, simulations are run, information is gathered, software is deployed and decisions are prepared. The agents may begin as tools used by individual scientists, then become more persistent co-workers, followed by project agents and organisational systems capable of coordinating work across a portfolio. At every layer, the difficult questions will concern context, authority and supervision as much as scientific intelligence.

Whether that produces better science will depend on more than the volume of work completed. We will need to ask whether agents improve the decisions that matter after supervision, correction and integration are counted. We will also need ways to retain the judgment currently held by experienced scientists without turning it into unquestionable fact, expose rather than hide uncertainty and prevent the efficiency of a swarm from outrunning its oversight. Finally, we will need to decide who benefits from the additional scientific capacity that these systems create.

I do not know exactly where this leads. But I am increasingly convinced that asking whether AI can make a sufficiently accurate scientific prediction captures only part of what is happening. It may be through the less glamorous work around that prediction—the gathering of context, operation of tools, coordination of people and allocation of effort—that AI changes science first.

References and further reading