”You cannot fail 1% of the time with patients.”

Manuel Corpas, Senior Lecturer at the University of Westminster and founder of ClawBio, joined our Head of EMEA, Tiffany Wang, on the latest episode of What The FLock to explain why the path from agentic AI to the clinic runs through validation, traceability, and reproducibility, not model capability alone.

That's the idea behind ClawBio, an open library of bioinformatics skills for AI agents. What began as the winning project at the UK AI Agent Hackathon × OpenClaw at Imperial College London is now used by researchers across three continents, with 1,017+ GitHub stars, 43 contributors, 90 skills, and recognition in Nature (Vol. 653, May 2026).

Frontier models are powerful, but biology operates on a different standard. A plausible answer isn't enough. It has to be checked against a validated method, every single time.

Rather than relying on a model's pre-trained knowledge alone, ClawBio packages bioinformatics workflows into reusable ”skills” that combine code, curated reference data, and defined procedures. Every answer follows a validated workflow instead of being generated from memory alone.

That creates three properties clinical AI depends on:

• Validation: Every result is checked against trusted methods and reference data, narrowing the gap between plausible and correct.

• Traceability: Every output can be traced back to the exact code, data, and workflow that produced it, making results auditable.

• Reproducibility: The same inputs produce the same outputs, allowing clinicians, researchers, and regulators to reproduce results with confidence.

In one early gene-drug comparison, Manuel shares that accuracy improved from roughly 75–80% to 100% after the task was wrapped in a bioinformatics skill. The model itself didn't become smarter, the workflow became more reliable by stopping relying on the model's memory and grounding every answer in a reproducible process.

The key takeaway: Domain-specific AI, especially in biology, won't be built by making models bigger alone. It will be built by creating systems that can validate every answer, show exactly how it was produced, and reproduce the same result every time.

Where do you think the reliability bar should sit for AI in genomics? Watch the full conversation.

Worth your time:

03:20–05:03 Why frontier models fall short in biology.

08:29–10:34 Validation, traceability, and reproducibility matter more than model capability in clinical AI.

11:45–15:50 How skill-guided workflows improved an early gene–drug comparison from ~75–80% accuracy to 100%.

15:50–18:56 17,000 evaluations across 16 models: what rigorous AI validation looks like in real-world science.

23:45–25:02 A candid discussion on private and local AI: what local processing can and cannot guarantee.

28:27–31:00 An open, transparent ecosystem is essential to democratizing agentic AI.

——

Guest info:

• Clawbio: https://github.com/ClawBio/ClawBio

• LinkedIn: https://www.linkedin.com/in/manuelcorpas

• X: https://x.com/manuelcorpas

• Website: https://manuelcorpas.com/

Follow FLock.io on:

• LinkedIn: https://www.linkedin.com/company/flock-io/

• X: https://x.com/flock_io

Visit our website: www.flock.io