Author(s): Maureen Doyle-Spare

Originally published on Towards AI.

Stanford’s 37,000 AI Agents Put Reasoning-Layer Governance on the Life Sciences Agenda

The scientific promise is extraordinary. As autonomous agents begin working across drug development, life sciences will need a way to preserve authorized meaning without giving up the speed and reach this new scale can create.

A glimpse of what agentic science can become

A Stanford Medicine research team has built a virtual biotech company with roughly 37,000 AI agents trained to support the full drug-development pipeline. Its organization mirrors an established biotech, with a chief science officer agent and specialized divisions working in parallel on activities ranging from identifying molecular targets to designing clinical trials.

The Stanford work is more than a demonstration of a single capable model. It shows how a large population of specialized agents might divide, coordinate and carry forward scientific work across a development program. Stanford reports that the agents analyzed and catalogued about 50,000 clinical trials in less than a week, work that the researchers said would have taken humans years. They identified biological characteristics associated with better clinical outcomes and used the wider virtual organization to investigate a lung-cancer target and design an antibody-drug conjugate strategy. A pharmaceutical company later independently pursued the same strategy. The Stanford paper was published in Science on September 17.

There is an important positive story here for biotechnology and clinical development. Agentic AI can expand the amount of scientific evidence that can be examined, allow specialized work to proceed in parallel and surface connections that would be extremely difficult for human teams to find at the same speed. In an industry where development programs can take years and clinical trials can cost tens or hundreds of millions of dollars, that capacity matters.

Stanford is appropriately clear about the boundary. Humans, physical experimentation and validation remain necessary before AI-generated findings become real-world scientific or clinical outcomes. The virtual biotech is a research demonstration, not a GxP deployment. That boundary matters, and it gives life sciences a concrete view of the type of autonomous scientific workflow that may eventually move closer to regulated decisions.

At 37,000 agents, governance changes

At this scale, governance cannot stop at asking whether each model produced an accurate answer. Agents retrieve evidence, apply working definitions, reconcile sources, coordinate with other agents and carry conclusions into later decisions. The important issue is the meaning created along the way. An agent works from the information, constraints, tools, prior decisions and institutional material available to it, then resolves what the task means in that particular context.

In my Agentic Workflow Drift in Life Sciences research, I call that resolved meaning the Operational Interpretation. It is the working meaning the agent has formed for the decision in front of it. That interpretation can be reasonable and still fall outside the meaning the organization intended to authorize. An agent may consult valid records, follow approved procedures and use source-backed information, yet combine those materials into an operational conclusion no controlled process explicitly approved. Immediately before execution, the interpretation can still be represented, evaluated and governed.

The distinction becomes more consequential as agent populations grow. Runtime Semantic Divergence describes a single decision made under meaning that departs from the institution’s authorized reference. Agentic Workflow Drift is the pattern that can form when such divergences recur without governance. The Stanford study does not report either condition, and there is no basis to suggest that it does. Its significance is that it makes the future control problem visible: autonomous interpretation can be distributed across thousands of specialized agents, with one agent’s conclusion becoming another agent’s starting point.

When every surrounding control looks clean

Life sciences already has one of the most disciplined control environments of any industry. Validation, data integrity, electronic records, quality systems, audit trails and qualified review remain indispensable. My AWD-LS research treats those disciplines as essential while identifying an additional governable object: the operational meaning an autonomous system resolves at runtime. Existing controls were not originally designed to supervise that layer.

This creates what I describe as the Invisible Failure. The source information can remain valid. The validated systems can operate correctly. Records can remain attributable, legible, contemporaneous, original and accurate. The audit trail can be complete, and the output can look plausible. Yet the Operational Interpretation assembled across those sources can still differ from the meaning the organization authorized. No conventional control needs to fail. The divergence sits between valid inputs and consequential action, in the meaning the agent has assembled from them.

The issue becomes especially important in life sciences because consequential terms are often dependent on context. Whether a deviation is critical, an adverse event is serious or expected, or a clinical subject is eligible may depend on the product, protocol, procedure, version, jurisdiction and applicable governing authority. A practical governance model must distinguish legitimate contextual variation from a departure in meaning that requires intervention. I describe that principle as Authorized Ambiguity: an organization may authorize more than one valid application of a term across defined circumstances, provided those variations remain within an approved institutional boundary.

What the institution has to preserve

A reasoning-layer control begins with a reference point that belongs to the institution. For any consequential decision, the organization needs an approved, versioned definition of what the relevant term means in that context, whether the decision concerns a critical deviation, a serious event, an expected adverse reaction or subject eligibility.

I call that authorized reference the Reasoning Baseline. It is the institution’s governed meaning for a defined decision class, product, protocol, procedure, jurisdiction or version, rather than a generic dictionary definition. The agent forms its working interpretation from the material available at runtime. The Semantic Control Plane represents the evidence of that interpretation as a Runtime Semantic State and compares it with the Reasoning Baseline before the agent moves into execution. This does not require access to private chain-of-thought. It requires enough evidence of the resolved meaning to determine whether it remains within the institution’s authorized range.

The Semantic Deviation Index (SDI) makes that comparison measurable. It is a pre-execution measure of divergence between the Runtime Semantic State and the Reasoning Baseline. It is distinct from model drift, explainability scoring and retrospective monitoring because its purpose is to surface material Runtime Semantic Divergence while the institution still has authority to intervene.

Governance that does not erase the scale

A scientific organization of 37,000 agents makes one operational fact difficult to avoid. Human review cannot grow one-for-one with every autonomous decision without removing much of the capacity these systems are intended to create. The alternative is selective, evidence-based governance that distinguishes ordinary authorized variation from the decisions that warrant intervention.

The Runtime Governance Applicability Test (RGAT) therefore belongs early in the deployment conversation. Before making a runtime-governance claim, the institution must establish that the decision class and deployment architecture expose enough operational state for a meaningful pre-execution evaluation. Some architectures will expose enough state, while others will not. The requirement is access to governable operational representations, not private chain-of-thought.

For decisions that qualify, the architecture separates measurement, verification and enforcement. SDI supplies the divergence signal, Verification of Runtime Semantic Resolution (V-RSR) evaluates whether the resolved state remains within institutionally authorized meaning, and the Deterministic Gate produces the binding Semantic Authorization disposition. Keeping those activities distinct prevents a measurement from being mistaken for an authorization decision and gives the institution a defined point at which it can permit, flag, hold or escalate a consequential action.

If the decision is authorized, Authorization Emission is the runtime event at which Semantic Authorization crosses the trust boundary into execution. Execution Authority exists downstream of that event. The sequence determines whether the organization can still intervene. Once authorization has crossed into execution, the task becomes reconstruction and remediation rather than prevention.

Life sciences should want the scale Stanford is demonstrating

Stanford has made the promise of agentic science easier to see. A virtual organization of tens of thousands of specialized agents can examine evidence at a volume and speed no conventional research team can match. Drug discovery, translational research and clinical development contain more evidence than any individual scientist or team can exhaustively reconcile. Life sciences should want the capacity to examine more of it, connect it faster and explore more hypotheses than conventional operating models permit.

The governance requirement follows from the same scale. When an agent’s interpretation can shape the next scientific analysis, influence another agent’s work or eventually enter a regulated workflow, the organization needs to preserve the difference between information that is valid and meaning that is authorized. It also needs to preserve legitimate Authorized Ambiguity rather than forcing complex scientific and regulatory judgment into a falsely rigid definition.

Runtime governance does not require a human to approve every action. It requires a way to identify the decisions where an agent’s resolved meaning falls outside the range the organization is prepared to authorize, and to intervene before that meaning becomes consequential action. The Reasoning Baseline preserves the institution’s authority, while the Semantic Control Plane makes divergence visible, verifies the resolved interpretation and establishes a binding intervention point before execution authority is emitted.

The Stanford demonstration shows what agentic scale may make possible for drug discovery and development. The next challenge is to ensure that scientific work can move at that scale without allowing ungoverned interpretations to become the operating standard. That is the role of reasoning-layer governance: to make responsible scale possible while preserving the speed and scientific reach that autonomous systems can create.

Sources and research

Stanford Medicine, “Virtual biotech company puts thousands of AI scientist agents to work on drug discovery,” September 17, 2026. Doyle-Spare, M.

Agentic Workflow Drift in Life Sciences: Extending the Reasoning-Layer

Risk Taxonomy to GxP-Regulated Pharmaceutical and Biotechnology Operations. Doyle-Spare, M. Agentic AI Drift Measurement and the

Semantic Deviation Index. Doyle-Spare, M. Agentic AI Systems

Governance: A Runtime Reference Architecture for the Reasoning Layer and the Semantic Control Plane in Regulated Financial Institutions. Doyle-Spare Research Library.

INTELLECTUAL PROPERTY: Maureen Doyle-Spare © 2026 Maureen Doyle-Spare. All rights reserved. No part of this article, including its text, figures, tables, and images, may be reproduced or distributed without the author’s prior written permission.

About the Author

Maureen Doyle-Spare is an enterprise governance architect and researcher in AI governance, specializing in agentic AI oversight and cybersecurity for autonomous and multi-agent systems. She is the originator of an agentic AI governance taxonomy built on a set of named constructs, developed for the layer above conventional AI controls, where autonomous agents interpret business meaning across fragmented enterprise systems and commit to it before acting.

Her research develops a runtime governance architecture and a foundational risk and threat taxonomy for the reasoning layer, the pre-execution stage at which an agent’s Operational Interpretation is formed and evaluated. The named constructs include Agentic Workflow Drift and Subversion (AWD/AWS), the Semantic Layer Integrity Attack (SLIA) as a cyber threat class targeting the reasoning-layer attack surface, the Semantic Control Plane (SCP) as a runtime governance mechanism, and the Semantic Deviation Index (SDI) as a measurement standard for semantic divergence. Her central thesis is that agentic systems do not fail the way traditional models fail: an agent can execute flawless steps against a meaning no institution authorized.

Her work spans behavioral visibility, pre-execution oversight, adversarial resilience, autonomous systems evaluation, and safe deployment, and policy crosswalks against the NIST AI RMF, MITRE ATLAS, STRIDE, ISO/IEC 42001, and the EU AI Act. She has provided public comments across multiple NIST AI initiatives, including the AI RMF, CAISI, COSAiS, NCCoE, and TEVV. Her working papers are available on SSRN and Zenodo.

Capstone reference architecture: https://doi.org/10.5281/zenodo.20749051
The Agentic AI Governence Playbook: https://www.maureendoylespare.com/
ORCID: https://orcid.org/0009-0009-6655-1394
SSRN Author Page: https://papers.ssrn.com/Sol3/Cf_Dev/AbsByAuth.cfm?per_id=10836296
ResearchGate: https://www.researchgate.net/profile/Maureen-Doyle-Spare/research
LinkedIn: https://www.linkedin.com/in/maureendoylespare/
GitHub: https://github.com/maureendoylespare/maureendoylespare
Substack: https://maureendoylespare.substack.com/

Originally published at https://maureendoylespare.substack.com on September 18, 2026.

Canonically published Stanford’s 37,000 AI Agents and Reasoning-Layer Governance | Maureen Doyle-Spare

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI