Myriad: a 10,000-node biotech agent skill graph

Myriad is a machine-readable map of biotech and pharmaceutical work: 100 domains, 1,000 workstreams and 10,000 atomic task nodes that agents can search and route to. It is a taxonomy, not a set of validated protocols: three nodes are implemented as complete skills.

npx kalaris-myriad stats

Maturity

Level	Count	Meaning
Taxonomy-defined nodes	10,000	Bounded tasks with routing metadata, objectives, provenance expectations, review boundaries and completion evidence
Implemented seed skills	3	Complete `SKILL.md` implementations with schemas, tool contracts, validation logic and examples

The three seed skills are diploid PacBio HiFi assembly optimization, variant-aware CRISPR off-target scoring, and HTS normalization, plate QC and hit calling.

Myriad does not authorize autonomous wet-lab execution, clinical decisions, GMP disposition, regulatory determinations or safety overrides. Operational use requires current primary sources, validated local procedures, appropriate authorization and qualified human review.

Coverage

Molecular and omics, structure and discovery, therapeutics and development, and manufacturing and translation: from genome assembly and single-cell transcriptomics through docking, ADME and CRISPR editing to bioreactor control, clinical trials, pharmacovigilance and GxP.

CLI

Command	Purpose
`kalaris-myriad stats`	Show release counts and maturity
`kalaris-myriad list domains`	List all 100 domains
`kalaris-myriad list tasks --domain N`	Browse task nodes
`kalaris-myriad search <query>`	Search titles, objectives and routing names
`kalaris-myriad show <id-or-name>`	Inspect one task node
`kalaris-myriad export --format md --output <path>`	Export all or part of the graph as `jsonl`, `json` or `md`
`kalaris-myriad validate`	Validate the installed release
`kalaris-myriad init [directory]`	Initialize a working subset

Add --json to stats, list, show and validate for machine-readable output.

npx kalaris-myriad search "variant-aware off-target" --limit 5
npx kalaris-myriad show MYR-D051-T041

Design principles

Precise routing: task names say when a node should activate and what decision artifact it produces.
Progressive disclosure: agents start at the global index, narrow to a domain or workstream, and load a full node only when needed.
Evidence before confidence: a score or model output is never treated as experimental, clinical or regulatory proof.
Fail closed: nodes support pass, warning, fail and no-call outcomes.

Explore the graph in the browser at myriad-explorer.vercel.app. Code is Apache-2.0 and content CC BY 4.0. Source: KalarisLabs/Myriad.