ClawBio HackathonAI Agents for Health
Build agents for genomics, pharmacogenomics, and digital health
23 April 2026 · University of Westminster · 115 New Cavendish St, London
Agentic engineering: the new frontier
Biology has become a data-saturated science. The tools to analyse it have not kept up.
A single human genome produces three billion base pairs. Clinical interpretation requires alignment, variant calling, functional annotation, literature cross-referencing, and synthesis into actionable reports. Each step has its own software, dependencies, version history, and configuration.
Workflow managers (Nextflow, Snakemake, Galaxy) emerged because manual orchestration is unsustainable. But the barrier to entry remains high. A biologist who wants to analyse their own sequencing data must learn to code, hire someone who can, or rely on graphical interfaces that may not support the analysis they need.
The models are good enough. The harness is not.
Modern LLMs can write, debug, and execute code. They can plan multi-step operations, adapt based on intermediate results, and coordinate dozens of tools.
But a general-purpose LLM generating bioinformatics code from scratch is unreliable. Output varies between sessions. It lacks the specificity that domain experts build into workflows over years. It halluccinates gene-drug associations and invents variant classifications.
The problem is not the model. The problem is the harness: what constrains the model, what tools it can call, what guardrails prevent silent errors. Agentic engineering is the discipline of building that harness.
What is agentic AI?
The first wave of LLMs in the life sciences was information retrieval: summarising papers, answering questions about pathways, extracting structured data from text. Useful but incremental.
The second wave is qualitatively different. When connected to file systems, databases, and command-line tools, LLMs become autonomous agents that plan multi-step operations, execute them, and adapt based on intermediate results.
The shift: from AI that tells you things to AI that does things. The researcher's role shifts from constructing the analysis to evaluating it. From production to judgement.
What is an agent?
Agentic genomics is the use of autonomous AI agents, powered by large language models and operating within domain-constrained skill libraries, to discover, plan, execute, and iteratively refine multi-step genomic analyses, where the agent exercises runtime decision-making over tool selection, parameterisation, error handling, and output evaluation.
Corpas, Fatumo, Guio. "Agentic Genomics: From Pipeline Automation to Autonomous Validation." Cell Genomics (submitted).
Four necessary conditions: autonomy (runtime decisions), domain constraint (skill libraries, not ad hoc code), iterative refinement (error diagnosis and self-repair), and natural language mediation (no programming required).
How does it work?
Traditional workflow
- Researcher writes code
- Configures tools manually
- Runs pipeline
- Interprets results
Bottleneck: code production
Agentic workflow
- Researcher describes intent in natural language
- Agent discovers and executes skills from a library
- Agent handles errors and adapts
- Researcher validates results
Bottleneck: validation and judgement
What agentic genomics is not
Not workflow automation
Nextflow and Snakemake execute a fixed DAG specified by a human. The workflow manager does not decide at runtime which tools to use. Agentic genomics builds on workflow infrastructure but is not reducible to it.
Not AutoML
Systems that search over hyperparameters operate within a fixed search space defined by a human engineer. They optimise within constraints; they do not formulate the analysis strategy.
Not LLM-assisted scripting
Using a chatbot to generate a Python script for variant filtering is useful but is not agentic. The LLM produces code; the human executes it. Sometimes called "vibe coding," it is a precursor but lacks the defining properties.
Not a chatbot
Biomedical AI copilots that answer clinical questions operate as information retrieval systems. They do not execute multi-step computational analyses against real data.
Why is it useful?
| Metric | Human-directed | Agent-mediated |
|---|---|---|
| Setup time | 2-4 hours | 5-15 minutes |
| Monitoring required | Continuous | Minimal (agent handles errors) |
| Reproducibility | Variable | High (skill specification fixed) |
| Error recovery | Manual debugging | Automated diagnosis and repair |
| Prerequisite expertise | Bioinformatics training | Domain knowledge for validation |
| Primary failure mode | Config errors, version conflicts | Silent plausible-looking errors |
Based on exome and scRNA-seq comparisons across multiple independent systems (Corpas et al., Cell Genomics).
How does this apply to genomics?
An emerging ecosystem
Several independent groups have converged on architectures that share the core properties of agentic genomics:
CellAtria (AstraZeneca)
Agentic framework for single-cell RNA-seq. Dialogue-driven, document-to-analysis automation. Prioritises reproducibility over flexibility.
AutoBA
Autonomous agent for multi-omic analyses. Users supply minimal input; agent plans, generates, executes, and self-repairs code.
Bio-Copilot
Multi-agent system with self-reflection protocols, shared knowledge database, and structured human-agent collaboration.
ClawBio
Community-driven skill library. Domain experts encode knowledge into agent-executable modules without needing to become software engineers.
This convergence from independent groups, using different architectures but arriving at similar design principles, is evidence that agentic genomics reflects a genuine structural shift.
What is ClawBio?
An open-source toolkit of AI agent skills for genomic analysis.
A skill is a self-contained, versioned unit of bioinformatics functionality: a SKILL.md contract that encapsulates code, configuration, data references, I/O specifications, and test suites. Skills are designed to be discovered and executed by AI agents through semantic search and natural language invocation.
How can it help me do better research?
Variant interpretation
VEP, ClinVar, gnomAD, ACMG classification, pharmacogenomics (CPIC). One command, structured report.
GWAS + PRS
Query 9 databases in parallel, compute polygenic risk scores, fine-map loci with SuSiE. All from summary statistics.
scRNA-seq
QC, doublet removal, clustering, marker genes. Scanpy pipeline wrapped in a single skill with demo data.
Equity audits
HEIM equity scorer measures how well a dataset represents diverse populations. Flags ancestry bias in analyses.
Literature + databases
PubMed summariser, UKB Navigator, Galaxy Bridge, protocols.io. Connect your analysis to the knowledge layer.
Clinical reports
WES clinical reports (English and Spanish), profile reports. From raw variants to PDF, with disclaimers and provenance.
What are the limitations?
Agentic genomics lowers the barrier to generating analyses. It does not lower the barrier to evaluating them.
Silent plausible errors
The primary failure mode across all agentic systems: results that look correct but are not. One skill returned "all normal" pharmacogenomics for an empty input file.
Hallucination
Agents can cite non-existent gene-disease associations, fabricate references, or generate variant annotations that conflate unrelated loci. Skill constraints reduce but do not eliminate this.
Equity bias
Agents default to European-ancestry resources unless explicitly constrained. 86% of GWAS data is European. Agentic genomics can automate existing biases at unprecedented scale.
Validation gap
Domain expertise, familiarity with common failure modes, intuition about plausible results: these are built over years and cannot be shortcut by AI. A novice user will not catch the errors.
The democratisation is real, but partial. It expands the capacity to produce; it does not expand the capacity to judge.
How do I get started?
Today. Right now. In this room.
Challenge tracks
Track 1: New skill
Add a bioinformatics skill to ClawBio. Variant annotation, pathway analysis, clinical reporting, or your own idea. Must conform to SKILL.md template with demo data and tests.
Track 2: Agent workflows
Chain multiple skills to solve a multi-step genomics question. Produce something no single skill could produce alone.
Track 3: Equity hack
Use the HEIM equity framework to audit representation gaps. Score a GWAS, compare PRS accuracy across populations, or build an equity dashboard.
Schedule
| Time | Activity |
|---|---|
| 12:00 | Arrival and setup |
| 12:30 | Introduction to ClawBio and challenge briefing |
| 13:00 | Team formation and hacking begins |
| 16:30 | Demos and judging |
| 17:30 | Prizes and networking |
| 19:00 | Close |
Team formation: We will pair domain experts (biologists, clinicians) with developers. No bioinformatics experience required. Teams of 2-4.
Setup
git clone https://github.com/ClawBio/ClawBio.git
cd ClawBio
pip install -r requirements.txt
python skills/bio-orchestrator/orchestrator.py --list-skillsJudging criteria
| Criterion | Weight |
|---|---|
Does it work? --demo runs without errors | 30% |
| Real-world impact. Solves a genuine problem | 25% |
| SKILL.md quality. Complete, follows template, gotchas documented | 20% |
| Tests. At least one test file, edge cases considered | 15% |
| Presentation. Clear 3-minute demo | 10% |
How to submit
- Fork ClawBio/ClawBio on GitHub
- Create a branch named
hackathon/your-team-name - Add your skill under
skills/your-skill-name/ - Open a PR before 16:30 with a description of what it does
- Demo live in 3 minutes: show the problem, run the skill, show the output
Prizes: Best new skill, best agent workflow, best equity hack. All winning skills will be merged into ClawBio with full attribution.
Join the community
Let's build.