## Is AI Agent Architecture Built for Biology — or Just Borrowed from Software?

The core constraint is physical: wet-lab experiments take days to weeks, not milliseconds. That single fact separates AI-assisted drug discovery from AI-assisted software development — and it is the central argument that [Cradle](https://synbiointel.com/companies/cradle) CEO and co-founder Stef van Grieken made in a July 2026 interview with R&D World. With Google DeepMind, Microsoft, Amazon, NVIDIA, OpenAI, and Anthropic all having launched or upgraded science-focused AI products within roughly 18 months, the pressure to transpose the coding-agent playbook onto life sciences has never been higher. Van Grieken's position: the analogy is structurally flawed, and the industry is only beginning to understand why. Deloitte pegged the average cost of bringing a drug from discovery to launch at $2.67 billion in 2025, a figure that reflects a decade-plus timeline no software harness can compress on its own.

---

## The Coding-Agent Playbook and Its Biological Limits

The momentum behind AI science tools is real and financially substantial. Anthropic reported its company-wide annualized revenue run rate crossed $47 billion in May 2026, up from $14 billion in February, with Claude Code's run-rate revenue doubling since early 2026 to more than $2.5 billion in May. OpenAI said Codex surpassed 5 million weekly active users by early June, up more than sixfold since the desktop app launched in February. SpaceXAI — formerly xAI, now the generative AI arm of SpaceX — agreed to acquire Cursor for $60 billion.

Frontier labs are now attempting to redirect that momentum toward biology. The underlying architecture is largely the same: a foundation model supplies the reasoning layer, while a surrounding software harness grants the agent access to files, databases, and external tools, then allows it to run those tools and respond to the results.

"The code is there and it works or it doesn't," van Grieken said. In software, an agent can execute, inspect an error, and revise within minutes. In biology, that loop involves physical synthesis, transformation, culture, assay, and analysis — a cycle that can take weeks and costs real consumables.

"Getting a drug, or any biomolecule, to market involves a long list of things you have to do to get somewhere," he added.

---

## Where Current AI Science Tools Actually Deliver

Van Grieken draws a sharp line between what current AI science platforms do well and where they fall short. The areas of genuine near-term value, in his framing, are document-heavy and data-rich: literature synthesis, hypothesis generation, competitive landscape analysis, and clinical operations.

The numbers from pharma deployments support that. AWS reported in 2025 that Novo Nordisk's work with AWS, MongoDB, and Anthropic cut the time required to generate clinical study reports by more than 90% — a task that previously occupied teams of writers for weeks.

"That's core, chatbot functionality," van Grieken said, referring to the current generation of frontier lab science products. IQVIA has partnered with NVIDIA to develop agentic services automating tasks, connecting market data, and accelerating clinical trials. Salesforce and Veeva are pushing agents into biopharma's operating layer. These are meaningful efficiency gains, but they sit upstream of the wet bench.

OpenAI trained GPT-Rosalind, released in April 2026 as a research preview, specifically for biology, drug discovery, and translational medicine. Anthropic and Google DeepMind build their science offerings on existing foundation models. Despite those architectural differences, van Grieken's analysis groups them together: sophisticated copilots for knowledge work, not autonomous molecular engineers.

---

## The Computational Layer: What an Agent Can Orchestrate

Van Grieken does credit the harness-based architecture with genuine utility at the computational tier. An agent equipped with the right tools can call [AlphaFold](https://synbiointel.com/glossary/alphafold) 3 or ESMFold2 to predict protein complexes, run protein-design pipelines such as BoltzProt-1 to generate and rank prospective binders, or orchestrate mechanistic disease models built in MathWorks' MATLAB-based SimBiology or implemented with ODE solvers in Fortran or Python — varying parameters across thousands of runs and surfacing a narrower set of perturbations for a scientist to examine.

That workbench can connect to PubMed, Semantic Scholar, arXiv, and FDA resources, running established bioinformatics tools such as bcftools against locally stored whole-genome sequencing files. "You can take some of the tooling people have used in bioinformatics and use a harness to run it," van Grieken said.

The caveat is that scientists must still verify citations and determine whether the resulting analysis is scientifically appropriate. Computational cycles are fast and cheap; they do not substitute for experimental validation.

---

## Cradle's Position in the Stack — and Why Experimental Data Is the Scarce Asset

[Cradle](https://synbiointel.com/companies/cradle) — headquartered in Amsterdam with a lab that houses automated liquid-handling and analytical equipment — sits specifically at the molecular engineering layer. Its software models the relationship between protein sequences and desired functional properties, then generates and ranks new sequences for customers to test in their own labs. Customers retain the resulting experimental assets; Cradle sells software, not drug pipelines.

That model puts Cradle in direct contrast with vertically integrated players. [Generate:Biomedicines](https://synbiointel.com/companies/generate-biomedicines) and [Absci](https://synbiointel.com/companies/absci) combine AI software, wet labs, and internal drug pipelines. Generate has pushed an AI-designed anti-TSLP antibody into Phase 3. [EvolutionaryScale](https://synbiointel.com/companies/evolutionaryscale) supplies protein foundation models; Chai Discovery focuses on de novo antibody design; Boltz and Tamarind Bio provide inference tools for structure prediction and related tasks.

Van Grieken's dependence on experimentally measured customer data points to a broader structural issue he raises: biology lacks a shared, continuously expanding corpus comparable to GitHub. The implication — not fully developed in the source text — is that proprietary wet-lab data becomes a durable competitive moat in a way that publicly scraped training data does not.

"There are many problems in biology where text or images or coding is not going to help you as much," van Grieken said.

---

## Industry Trajectory: The Feedback-Loop Problem Will Sort the Field

The hyperscaler rush into biology is real, well-funded, and not slowing. But van Grieken's analysis, grounded in years of operating at the protein-design layer, identifies the structural reason the market will likely segment: tasks with fast, legible, digital feedback loops will see rapid AI automation; tasks requiring physical iteration in a wet lab will remain expensive, slow, and human-intensive for longer than the current hype cycle implies.

That creates at least two distinct market layers. The first — document processing, clinical operations, literature synthesis — is already being commoditized by general-purpose frontier models. The second — [computational protein design](https://synbiointel.com/glossary/computational-protein-design), sequence-to-function modeling, experimental design for novel molecules — requires specialized training data, domain-specific architectures, and tight integration with wet-lab workflows. Companies that can close that loop, whether by owning lab infrastructure like Generate and Absci or by embedding deeply in customer workflows like Cradle, hold a structural advantage that no foundation model upgrade alone will dissolve.

---

## Key Takeaways

- **The coding-agent analogy breaks in biology.** Software agents iterate in minutes on digital feedback; wet-lab experiments take days to weeks and carry real cost per cycle.
- **Deloitte estimated average drug-development cost at $2.67 billion in 2025** — a figure reflecting the irreducible physical complexity of the discovery-to-launch pipeline.
- **Current AI science tools excel at document-heavy tasks.** Novo Nordisk's AWS/Anthropic deployment cut clinical study report generation time by more than 90%.
- **Anthropic's annualized revenue run rate crossed $47 billion in May 2026**, up from $14 billion in February, with Claude Code's run-rate revenue exceeding $2.5 billion — illustrating the commercial scale driving the biology push.
- **Cradle targets the molecular engineering layer** specifically: sequence generation and ranking for customer labs, not internal drug pipelines.
- **Biology lacks a GitHub equivalent.** Proprietary experimental data — not publicly scraped corpora — is the scarce training asset for protein-design models.
- **The market will segment:** general-purpose frontier models will commoditize knowledge-work tasks; specialized models with experimental data pipelines will dominate molecular design.

---

## Frequently Asked Questions

**Why can't AI coding agents work the same way in drug discovery?**
AI coding agents succeed because they can execute code, receive an error, and iterate within minutes — all in a digital environment. Drug discovery requires physical synthesis and wet-lab testing, a feedback loop that takes days to weeks and costs real materials per cycle. Stef van Grieken of Cradle frames this as a structural difference, not a temporary technical gap.

**What is Cradle's business model and how does it differ from Generate:Biomedicines or Absci?**
Cradle sells protein-engineering software to customers who run experiments in their own labs and retain the resulting assets. Generate:Biomedicines and Absci are vertically integrated, combining AI software with internal wet labs and proprietary drug pipelines. Generate has an AI-designed anti-TSLP antibody in Phase 3.

**What can AI science agents actually do well in biopharma today?**
The clearest gains are in document-intensive workflows: literature synthesis, hypothesis generation, clinical study report drafting, and data analysis. Novo Nordisk reportedly cut clinical documentation time by more than 90% using a workflow with AWS, MongoDB, and Anthropic.

**What is GPT-Rosalind and how does it differ from other AI science tools?**
OpenAI released GPT-Rosalind in April 2026 as a research preview trained specifically for biology, drug discovery, and translational medicine. Anthropic and Google DeepMind build their science offerings on existing foundation models rather than training biology-specific ones from scratch.

**Why does biology lack a training data corpus comparable to GitHub?**
Code is produced, versioned, and shared publicly at massive scale. Wet-lab experimental data — measurements of how specific sequences perform against specific assays — is generated slowly, is expensive to produce, and is almost never shared publicly. This makes proprietary experimental datasets a durable competitive asset for companies building protein-design AI, in contrast to the broadly available corpora that underpin general-purpose language models.