# Google proposes a three-layer agent stack for biotech

> Google's biotech reference design separates agent planning, GPU scientific computation and access controls. The combined workflow remains an unmeasured proposal.

By BIG CHANGE Editorial

Published: 2026-10-18T17:10:34.892Z
Updated: 2026-10-18T17:10:34.892Z
Canonical: https://bigchange.ai/blog/google-biotech-agent-architecture

![Conceptual charcoal illustration of a researcher examining a ball-and-stick molecular model, with generic compute racks in a separate room behind glass.](https://bigchange.ai/api/media/file/google-biotech-compute-concept-v2.png)
AI-generated conceptual illustration by BIG CHANGE; no actual Google facility, completed biotech workflow, drug candidate or test result is depicted.

Google Cloud has published a reference architecture for biotech agents that separates a planning agent from GPU-heavy scientific computation and the controls around both. The [October 16 design post](https://cloud.google.com/blog/topics/healthcare-life-sciences/architecting-scalable-agentic-biotech-workloads) describes how a drug-discovery team could coordinate protein folding, molecular generation and lead review on Google Cloud. It presents a pattern and an example stack, without a measured customer deployment, throughput result or validated drug candidate.

The distinction matters for teams evaluating agent systems in research. A language model can route work and summarize results, but a protein structure or docking trajectory is too large to treat as ordinary chat context. Scientific computation also has its own runtime, queue and access controls. Google's proposal puts those parts behind explicit boundaries so a team can inspect each one before relying on the combined workflow.

## Three layers, with different jobs

The first layer is a lead orchestrator on [Agent Runtime](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime). Google's example gives it two specialist subagents, one for lead optimization and another for safety screening. The lead agent decomposes a request, routes work and assembles an answer. In this proposed design, internal agent handoffs occur in process. Agent Runtime is a managed deployment environment; Google documents a project setup, Agent Platform SDK for Python, agent development and deployment before the agent can be called. A containerized agent must meet its [runtime contract](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime).

The second layer runs the expensive scientific jobs. Google's example places NVIDIA's ESMFold and GenMol inference microservices on Google Kubernetes Engine using G4 virtual machines with RTX PRO 6000 Blackwell GPUs. Protein folding and molecule generation are examples of work the orchestrator would request, rather than computations performed by the general-purpose language model. The post recommends keeping at least one GKE node active and loading model weights when pods start to reduce cold-start delay. A warm node and loaded weights consume resources; the post supplies no latency, utilization or cost comparison for that choice.

The third layer governs access. [Agent Identity](https://docs.cloud.google.com/iam/docs/agent-identity-overview) assigns an agent a SPIFFE identity and managed X.509 certificate, which can be used with IAM permissions and audit logs. [Agent Gateway](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/gateways/agent-gateway-overview) can enforce access rules on supported ingress and egress paths. [Agent Registry](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/agent-registry) catalogs agents, MCP servers, tools and endpoints. Google also proposes Model Armor inspection and human approval for consequential steps. These controls require deployment and policy configuration; their presence in an architecture drawing does not establish that a particular data path is private or that every decision is audited.

## Move files through storage, not through the prompt

The most transferable part of the design is its treatment of large scientific payloads. The proposed workflow writes PDB files, docking poses and trajectory logs to Cloud Storage. Agents exchange object references and short summaries, then retrieve the underlying data when a tool needs it. That avoids pushing coordinate arrays through an LLM context window and gives the team a separate place to control file access and retention.

Google describes signed object URIs in this pattern. A signed URL is itself an access credential for its stated scope and lifetime, so teams still need to set expiration, permissions and logging for the actual workflow. The design post does not document a complete storage policy, residency configuration or threat model for a production biotech pipeline. The claim that a privately hosted compute container keeps proprietary structures inside an enterprise perimeter depends on every component that receives or forwards those structures, including storage, model calls, logs and outbound endpoints.

The post also recommends wrapping a long simulation chain behind a clear tool interface and emitting progress at completed stages. Those are design choices to assess against the actual backend queue. Sending more requests in parallel will not help when a scientific service serializes them internally. The article offers no benchmark showing that its proposed configuration has deterministic throughput.

## Where to put the human decision

Google's proposal lets agents handle reversible preparation, such as looking up a structure or resolving a chemical name, while requiring a person to approve high-cost or consequential actions. Its examples include a multi-GPU docking run, moving a lead toward wet-lab synthesis and archiving a program. A research organization would need to define those gates, the approver and the evidence shown at each gate. The post does not report a validated safety-screening result or establish that a model can make pharmacological judgments without expert review.

The governance documentation narrows some of the blog's broad language. Gateway supports both ingress and egress with Agent Runtime, while its Gemini Enterprise support is egress only. Unregistered destinations can still be addressed when an explicit IAM policy targets their URL; registration itself is therefore not a universal block on unknown tools. Model Armor checks are among the guardrails a team *can configure*. Agent Identity's documented limitations include the inability to grant legacy Cloud Storage bucket roles to agent principals, and IAM bindings need manual cleanup when an agent is deleted.

For validation, Google Cloud's separate [four evaluation lessons from an agent plugin](https://cloud.google.com/blog/topics/developers-practitioners/four-eval-engineering-lessons/) are useful context, though that plugin is a different system. The team found that judging only an agent's final answer missed actions visible in tool-call logs. It recommends preserving concurrent call structure, comparing live and plan-only runs, and keeping expert-curated baseline tasks. Applied to the biotech proposal, that points to checking the actual storage writes, model endpoints, permission decisions, queue behavior and human approvals. BIG CHANGE has not built or tested this architecture.

Our earlier account of [Google Antigravity's antimicrobial peptide triage](https://bigchange.ai/blog/google-antigravity-antimicrobial-peptide-candidate-triage) concerns a different scientific-agent task. It does not establish that the infrastructure proposed here has been built or validated.

## Access, cost and release status

A team would need a Google Cloud project, the relevant APIs and IAM setup, Agent Runtime deployment, GKE GPU capacity, storage and configured governance components. Availability and limits depend on the selected regions, quota and product configuration. The [Agent Runtime documentation](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime) points to region and quota lists and says a free tier is available. The [Agent Platform pricing page](https://cloud.google.com/products/gemini-enterprise-agent-platform/pricing) charges for runtime compute and memory and describes Gateway egress billing; model usage, GPU infrastructure and Cloud Storage add their own costs. Google's [pricing calculator](https://cloud.google.com/products/calculator) can help estimate a chosen configuration. A continuously warm GPU pool should be included in that estimate.

The component documents describe usable services, but they do not turn the combined biotech pattern into a generally available, measured product. Registry's standalone skills are explicitly marked Preview, while the pages checked here do not give one blanket release-stage label for every element in the proposed stack. Compatibility among the precise model containers, network controls and research applications must be checked for the intended region and deployment. This article is a documentation-based examination of the October 16 proposal, not a setup guide or an account of a working installation.

## The big change

Google has laid out a way to divide a biotech agent workflow into planning, dedicated scientific compute and governed access, with large structures passed by reference and human approval reserved for costly or consequential actions. The design gives research infrastructure teams a concrete set of boundaries to evaluate. Its speed, isolation, scientific reliability and operating cost remain questions for a configured deployment and measured tests.

## Sources & further reading

- [Google Cloud, “Architecting autonomy”](https://cloud.google.com/blog/topics/healthcare-life-sciences/architecting-scalable-agentic-biotech-workloads): The October 16 reference architecture and its proposed components, workload patterns and approval gates.
- [Agent Runtime documentation](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime): Managed agent deployment, prerequisites, framework support, region and quota links, and pricing entry point.
- [Agent Identity overview](https://docs.cloud.google.com/iam/docs/agent-identity-overview): Identity and credential mechanics, IAM lifecycle and a Cloud Storage role limitation.
- [Agent Gateway overview](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/gateways/agent-gateway-overview): Enforcement model, runtime modes, configuration-dependent guardrails and stated limitations.
- [Agent Registry documentation](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/agent-registry): Registered resource types and the Preview status of standalone skills.
- [Agent Platform pricing](https://cloud.google.com/products/gemini-enterprise-agent-platform/pricing): Runtime and Gateway charging basis; estimate the full stack separately.
- [Google Cloud's four evaluation lessons](https://cloud.google.com/blog/topics/developers-practitioners/four-eval-engineering-lessons/): Context from another agent plugin on checking execution traces and evaluation design.

Sources checked October 18, 2026. BIG CHANGE did not deploy or benchmark the proposed biotech system.

## Sources

- [Google Cloud biotech agent architecture](https://cloud.google.com/blog/topics/healthcare-life-sciences/architecting-scalable-agentic-biotech-workloads) — October 16 design proposal and layer descriptions.
- [Agent Runtime](https://docs.cloud.google.com/gemini-enterprise-agent-platform/build/runtime) — Deployment workflow, supported frameworks, contract, quota and pricing links.
- [Agent Identity overview](https://docs.cloud.google.com/iam/docs/agent-identity-overview) — Identity credentials, IAM lifecycle, audit and limits.
- [Agent Gateway overview](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/gateways/agent-gateway-overview) — Modes, policies, unregistered destinations and limitations.
- [Agent Registry](https://docs.cloud.google.com/gemini-enterprise-agent-platform/govern/agent-registry) — Catalog and Preview label for standalone skills.
- [Agent Platform pricing](https://cloud.google.com/products/gemini-enterprise-agent-platform/pricing) — Runtime and Gateway charging basis.
- [Google Cloud pricing calculator](https://cloud.google.com/products/calculator) — Configuration-specific estimates.
- [Four evaluation engineering lessons](https://cloud.google.com/blog/topics/developers-practitioners/four-eval-engineering-lessons/) — Context from another plugin, not validation of the biotech design.