For builders, the working assumption has been one engineer, one agent — the mannequin Claude Code and comparable instruments. At VB Remodel 2026, James Zou, affiliate professor of biomedical information science at Stanford College, argued that assumption is about to interrupt: the subsequent frontier isn't a single, extra succesful agent, it's tens of 1000’s of them collaborating.
For builders and product builders, probably the most essential takeaway from Zou’s presentation is how these huge programs are orchestrated. His crew's analysis provides a sensible blueprint for connecting legacy databases to AI orchestration layers and designing environments that allow 1000’s of brokers to collaborate.
Emulating the group — the digital biotech
Zou’s challenge started as a "Digital Lab" consisting of 5 to eight brokers structured to reflect his bodily Stanford lab. The setup included an AI professor appearing because the principal investigator and AI college students with distinct specialties holding common group conferences.
"We additionally created for the brokers a duplicate of Stanford, an agent college, the place the brokers can really go to the college and do supervised fine-tuning to enhance their experience of their particular domains," Zou famous.
The digital lab efficiently designed new nanobody proteins for latest COVID variants.
"What is absolutely thrilling to us is that these AI-designed nanobody proteins really labored a lot better than the earlier human-designed nanobodies when it comes to binding to the latest totally different viruses," Zou stated.
Following this wet-lab validation, the crew expanded their ambition. They transitioned from emulating a single analysis crew to modeling an enormous company construction.
The ensuing system, dubbed the Digital Biotech, includes tens of 1000’s of specialised AI brokers overseen by a Chief Scientific Officer (CSO) agent. It operates by means of distinct company divisions, akin to goal discovery, molecule design, and scientific trials.
"Working with the CSO agent are totally different divisions that mirror the divisions present in a human biotech or pharma firm," Zou defined — one centered on figuring out drug targets, one other on designing molecules, a 3rd on security and scientific trials. Particular person brokers specialize additional inside a division, he stated. "Underneath the goal discovery division, we'll have one agent that makes a speciality of taking a look at all of the genetics information, one other agent that appears in any respect the genomics information and single-cell information, and so forth."
The multi-agent benefit
As basis fashions develop extra succesful, builders face a core architectural dilemma: Why distribute workloads throughout tens of 1000’s of specialised brokers as an alternative of channeling all computing assets right into a single, omniscient mannequin?
Zou's crew ran a head-to-head comparability of a multi-agent crew towards a single agent tasked with the identical scientific problem. The multi-agent ecosystem created friction and interplay that produced higher options that had been extra resilient towards compounding errors.
"In these scientific digital labs, the brokers really get into debates and disagreements. They must persuade the opposite AI scientists [of] their concepts, and all of that elicits way more inventive and strong reasoning in comparison with if in case you have a single mannequin making an attempt to do the issue by itself from scratch," Zou stated.
The orchestration bottleneck
When scaling to tens of 1000’s of brokers, orchestration turns into the first bottleneck. The system requires a unified context layer that enables brokers to synthesize data from numerous instruments, datasets, and historic information.
Many enterprise groups try to unravel information integration by wrapping current databases with an MCP. Nonetheless, legacy programs are usually not very pleasant to brokers. For example, dropping a PDF of a analysis paper into an agent's context window is inefficient, and customary textual content fashions battle to interpret complicated figures and tables, resulting in hallucinations.
"Even if you happen to wrap an MCP across the current databases and APIs, that doesn't remedy the underlying drawback: the interface and APIs are usually not appropriate for brokers," Zou stated. He added that current databases are designed to be consumed by people or pre-AI algorithms.
To resolve this, Zou's crew created Paperclip. The platform depends on a core energy of contemporary LLMs: their potential to put in writing code and navigate file programs. As a substitute of forcing brokers to question brittle, database-specific APIs, Paperclip digitizes unstructured information and maps disparate databases right into a unified, AI-native digital file system.
This construction permits brokers to entry data from tens of millions of papers utilizing customary file-system operations.
"This principally exhibits that we are able to get a lot better accuracy if you happen to use Paperclip, and we are able to cut back the time and the price by over an order of magnitude in comparison with if you happen to use brokers with out these AI-native scientific infrastructures," Zou said.
Actual-world validation
To check the sensible output of this structure, Digital Biotech spun up 37,000 "scientific trial brokers" to synthesize fragmented trial information. These brokers recognized single-cell options that predict trial success — drug targets supported by these options had been about 50% extra more likely to attain market than comparable medication with out them.
The system then autonomously designed an antibody-drug conjugate (ADC) focusing on the CD276 protein for lung most cancers. The brokers accomplished this design autonomously, relying completely on information printed previous to January 2025.
A number of months later, Zou stated, pharmaceutical firm Merck independently developed and validated the identical therapeutic design — which went on to obtain breakthrough designation from the FDA. He characterised this as "a third-party exterior validation of the therapeutic design offered by the digital biotech brokers."
Designing ecosystems, not workflows
As multi-agent programs scale, leaders should rethink how they handle these digital workforces. Zou advocated for shifting from designing inflexible workflows to creating open environments. Workflows dictate the precise steps an agent ought to take, much like managing a junior worker. Environments present the infrastructure, guardrails, and incentives for brokers to collaborate on open-ended issues.
"In workflows, we're making an attempt to inform brokers what to do and how you can do their job. However in environments, we're offering the infrastructures, the incentives, and the guardrails, however in any other case we go away it open to incentivize brokers to collaborate," Zou stated.
Optimization at scale means engineering the setting somewhat than fine-tuning particular person fashions. Whereas single brokers can enhance through reinforcement studying or supervised fine-tuning within the agent college, the success of an enormous multi-agent system depends on adjusting the parameters governing their collaboration.
"On the multi-agent [side], we're not really fine-tuning and altering the person fashions anymore, however we're optimizing the setting," Zou defined. "The setting itself is the thing that we optimize to enhance the brokers."
