AI Agents Are Becoming Teams Instead of Tools

The next AI breakthrough may not be a smarter agent

The first generation of AI coding tools established a simple mental model. One person gives an agent a task, the agent works through it, and the human reviews the result.

That model is already becoming familiar.

Stanford University researcher James Zou is proposing something much larger. Instead of asking whether one AI agent can become capable enough to solve a difficult scientific problem, his team is exploring what happens when tens of thousands of specialized agents operate together as an organization.

At VB Transform 2026, Zou described a system that evolved from a small virtual research laboratory into what his team calls a Virtual Biotech. The system includes specialized agents for target discovery, molecule design, safety and clinical trials, with a Chief Scientific Officer agent coordinating the broader organization.

The scale is the important part.

The system eventually involved 36,000 clinical trial agents, and one of its independently generated therapeutic designs was later developed and validated by Merck.

That does not mean AI has replaced pharmaceutical research. It does suggest that the architecture of AI systems may be moving from individual assistants toward something closer to digital organizations.

From a virtual laboratory to a virtual company

The original experiment was relatively small.

Zou's team created a Virtual Lab containing roughly five to eight agents designed to resemble the structure of a physical Stanford research laboratory. An AI professor acted as the principal investigator, while AI students performed specialized research roles and participated in group discussions.

The researchers even created a virtual version of Stanford where agents could receive supervised fine-tuning and improve their expertise in particular scientific domains.

This is an important architectural shift.

The objective was not simply to give one model more information. It was to create specialization, hierarchy, interaction and institutional memory.

The early system was able to design new nanobody proteins for recent COVID variants. According to Zou, the AI-designed proteins performed better in binding to the newer viruses than earlier human-designed nanobodies when tested in the laboratory.

The team then expanded the concept.

Instead of reproducing one research group, it attempted to reproduce the organizational structure of a biotechnology or pharmaceutical company.

That produced the Virtual Biotech.

What is different about the Virtual Biotech

A conventional AI system might have one model responsible for a complete research problem.

The Virtual Biotech takes the opposite approach.

Different agents have different responsibilities. One division focuses on discovering biological targets. Another focuses on molecule design. Another works on safety and clinical trials.

Even inside those divisions, agents specialize further.

For example, within target discovery, one agent might analyze genetics data while another works with genomics or single-cell datasets.

The architecture resembles a company because the problem itself resembles a company-sized problem.

Drug discovery is not one task. It is a chain of interdependent scientific decisions.

The advantage of specialization is therefore not simply that agents can perform more tasks. It is that they can challenge one another.

Zou's team compared multi-agent systems with a single agent working on the same scientific problem. The multi-agent system generated disagreements and debates between agents.

That friction matters.

If every agent immediately agrees with the first proposed answer, the system becomes an automated echo chamber. When specialized agents have to defend competing hypotheses, the interaction can generate more robust reasoning and reduce the risk that one early mistake propagates through the entire process.

In other words, the system does not only scale computation.

It scales intellectual diversity.

The real bottleneck is not intelligence

There is an obvious question.

If foundation models continue becoming more capable, why not simply build one extremely powerful agent?

Because intelligence is only one part of the system.

Once thousands of agents are working simultaneously, the problem becomes coordination.

What does each agent know? Which data can it access? How does it communicate with another agent? How does it interpret scientific documents? How does it avoid repeating work? How does the organization preserve useful discoveries?

This is where infrastructure becomes more important than model size.

Many companies currently attempt to connect existing databases to AI through interfaces such as MCP. But putting an interface around an old database does not automatically make that database useful to an AI agent.

Legacy databases were generally designed for humans or traditional software.

Scientific information creates an even harder problem.

A research paper may contain paragraphs, tables, diagrams, figures and experimental results. Simply placing the PDF into a model's context window does not guarantee that the model will correctly understand the underlying scientific information.

The problem is therefore not merely access to data.

It is the representation of data.

Paperclip is an example of AI native infrastructure

Zou's team developed a system called Paperclip to address this problem.

Instead of forcing agents to interact with a collection of complicated database-specific APIs, Paperclip converts unstructured scientific information into an AI-native virtual file system.

That matters because modern language models are particularly good at writing code and navigating files.

The system effectively gives agents a common environment for interacting with scientific knowledge.

According to Zou, this approach produced substantially better accuracy while reducing time and cost by more than an order of magnitude compared with agents operating without these AI-native scientific infrastructures.

The lesson extends beyond biotech.

Every enterprise has a similar problem.

Companies have decades of databases, documents, APIs, spreadsheets and internal applications. The AI agent may be extremely capable, but if the underlying information architecture was designed for humans rather than machines, the agent spends much of its time fighting the interface.

The next generation of enterprise AI infrastructure may therefore be less about building another chatbot and more about rebuilding the information layer for machines.

36,000 agents turn research into an organizational problem

The Virtual Biotech eventually deployed 36,000 clinical trial agents to synthesize fragmented clinical trial information.

The agents identified single-cell features associated with successful trials.

According to the research presented by Zou, drug targets supported by those features were approximately 50% more likely to reach the market than comparable drugs without those features.

That is a very different proposition from an AI writing assistant.

The system is attempting to identify patterns across enormous amounts of scientific evidence and connect those patterns to commercial outcomes.

The most striking demonstration came next.

The Virtual Biotech autonomously designed an antibody-drug conjugate targeting the CD276 protein for lung cancer, using information published before January 2025.

Several months later, Zou said that Merck independently developed and validated the same therapeutic design. The resulting therapy subsequently received FDA breakthrough designation.

The significance is not that an AI system magically invented a drug from nothing.

It is that a computational research organization operating from historical scientific information reached a therapeutic design that was independently pursued and validated by a major pharmaceutical company.

That is a meaningful external validation of the architecture.

The business model is bigger than software subscriptions

The economics could be even more important than the technology.

Traditional pharmaceutical research requires highly specialized employees, laboratory infrastructure, clinical operations and years of experimentation.

If AI agents can automate meaningful portions of target discovery, molecule design and clinical analysis, the economics of research could change.

The opportunity is not necessarily to eliminate scientists.

It is to increase the amount of scientific work that a relatively small number of humans can supervise.

That creates operating leverage.

A human researcher might supervise hundreds or thousands of computational experiments rather than manually executing each analytical step.

A pharmaceutical company could potentially explore more hypotheses before committing expensive laboratory resources.

The value proposition becomes straightforward.

Spend more computation to reduce wasted scientific experimentation.

That is potentially a very large market.

This is a platform problem, not merely a product problem

There is another important distinction.

A single AI agent is a product.

A system capable of coordinating thousands of agents is a platform.

The platform includes the models, data, tools, communication mechanisms, memory, evaluation systems, incentives and guardrails required for agents to operate together.

This changes where competitive advantage comes from.

The best model will not necessarily produce the best organization.

A slightly weaker model embedded inside a dramatically better environment may outperform a more powerful model operating inside a poor one.

That is why the orchestration layer becomes strategically important.

The winning companies may own the environment in which agents work rather than simply the agents themselves.

The competitors are not only AI companies

The obvious competitors are foundation-model companies and agent platforms.

But the deeper competitive field includes enterprise software, cloud infrastructure, data platforms and vertical software.

A pharmaceutical company with proprietary clinical data has an advantage that a general AI company cannot easily reproduce.

A cloud provider has the compute required to run thousands of agents.

A data company can provide structured access to specialized information.

A biotech company has scientific expertise and laboratory validation.

The strongest systems may combine all four.

This creates a familiar technology pattern. The model is important, but the surrounding ecosystem determines how much economic value can actually be extracted from it.

The biggest risk is coordination failure

Scaling from eight agents to 37,000 is not a simple multiplication problem.

More agents create more opportunities for conflicting conclusions, duplicated work, bad data and cascading errors.

There is also a verification problem.

An agent can produce a plausible scientific argument without producing a scientifically correct result.

That makes evaluation and human oversight critical.

The system therefore needs guardrails, incentives and mechanisms for resolving disagreement.

And there is another risk.

If the agents are optimized too aggressively toward a particular metric, they may discover shortcuts that improve the metric without improving the underlying scientific outcome.

At enterprise scale, this becomes an organizational design problem as much as a machine-learning problem.

Workflows are becoming environments

Perhaps the most important idea from Zou's presentation is the distinction between workflows and environments.

A workflow tells an agent what to do.

Step one. Search the database.

Step two. Analyze the document.

Step three. produce a report.

That approach works when the problem is predictable.

Scientific discovery is not predictable.

An environment is different. It gives agents infrastructure, incentives and guardrails while leaving room for agents to decide how to solve an open-ended problem.

This is closer to managing an organization than programming a workflow.

The implication is profound.

The optimization target changes.

Instead of continuously fine-tuning individual agents, developers increasingly need to optimize the environment in which those agents operate.

The question becomes less:

How do we make this agent smarter?

And more:

How do we create an environment where thousands of agents can collectively produce better outcomes?

What businesses should do now

Companies should not immediately deploy thousands of autonomous agents.

They should start by identifying where coordination creates economic value.

First, map existing data systems. Determine which databases, documents and APIs were designed for humans rather than AI.

Second, create an AI-native information layer. The objective should be to make corporate knowledge easy for agents to discover, interpret and reuse.

Third, introduce specialization. Instead of asking one agent to perform everything, create distinct roles for research, verification, analysis and execution.

Fourth, build evaluation into the environment. Every important agent output should have a measurable way to determine whether it was useful.

Fifth, keep humans in the loop where the cost of an incorrect decision is high.

Finally, optimize the system rather than obsessing over the individual model.

The model will change.

The environment is the infrastructure that compounds.

The larger implication

The first phase of enterprise AI was about productivity.

One employee gets an AI assistant.

The second phase is about automation.

An agent performs a workflow.

The emerging third phase is organizational.

Thousands of agents collaborate inside an environment designed specifically for machine labor.

Stanford's Virtual Biotech provides an unusually clear example because drug discovery is complicated enough that a single-agent architecture quickly reaches its limits.

The most interesting part of the experiment is therefore not simply the number 36,000.

It is the architecture behind that number.

The agents require specialized roles. They require shared information. They need a common interface to scientific knowledge. They need debate, evaluation, incentives and guardrails. And they need an environment in which collaboration produces better outcomes than isolated intelligence.

That points toward a future in which companies may not simply buy AI tools.

They may build AI organizations.

The competitive advantage will increasingly come from designing the environment in which those organizations operate.

And if that happens, the biggest AI opportunity may not be building a smarter employee.

It may be building the company that the employees happen to be.

15 SEO Keywords

AI agents, virtual biotech, Stanford AI, drug discovery AI, multi-agent systems, AI drug design, clinical trial AI, AI orchestration, AI infrastructure, Paperclip AI, pharmaceutical AI, autonomous agents, AI research lab, agentic AI, biotech automation

Conclusion

Stanford's experiment suggests that the next major step in AI is not necessarily a single model with unlimited intelligence.

It could be coordination.

Five agents become a laboratory. Thousands become a research organization. Tens of thousands begin to resemble an entire company.

That changes the architecture, the economics and eventually the competitive landscape.

The companies that understand this shift early will focus less on building isolated AI features and more on creating environments where specialized agents can access data, challenge one another, learn from outcomes and operate within carefully designed constraints.

The future of AI may therefore look less like one superhuman assistant sitting beside every employee.

It may look like an entire digital workforce operating behind the scenes.

And Stanford's 36,000-agent Virtual Biotech is one of the clearest demonstrations yet of what that future could look like.