CHORUS
Idle
Hierarchical Multi-Agent Virtual Cell

One agent per biological entity.
Molecule to tissue. Two Masters.

Chorus simulates biology from molecule to tissue with an intelligent agent at every level. Organ and organism levels are implemented and stay one request away; a default run stops at tissue, because that is the scale it can speak about. A Top-Down Master issues directives downward; a Bottom-Up Master verifies and aggregates upward — over a persistent, read-write organism context. Not incrementally better than prior systems. Orthogonal.

Top-Down · directive ↓Bottom-Up · aggregation ↑Homeostasis · verified
Grounded in the literature
Bunne 2024, CellWang 2025Virtual Lab, NatureCellForgeVCWorld 2025TxAgentGeneAgent

Live cascade

Top-Down Master

Sets directives from tissue → molecular

IDLE
🧬 Tissue
🕸️ Intercellular
🔬 Intracellular
⚛️ Molecular
Bottom-Up Master

Verifies + aggregates molecular → tissue

IDLE
Shared Organism Context
Awaiting patient profile…

Configure the organism on the left, then press Run Simulation to watch the cascade fire across the four cell-scale levels.

Nine limitations → nine mechanisms

No prior system has more than 4 of 9. Chorus has all 9.
Prior limitation

Close analogues cover only a subset of biological scales.

Hierarchy is a first-class primitive — six levels implemented from day one, four in a default run.

Prior limitation

Prior systems have one-way flow or an implicit scheduler — never a first-class coordinator pair.

Two symmetric Masters — Top-Down and Bottom-Up — expose bidirectional flow as an interface.

Prior limitation

Organism-wide variables are treated as external parameters, not runtime state.

A persistent, read-write organism context every agent consults and updates.

Prior limitation

Most multi-agent biology systems make agents that stand for scientists, not for biology.

Every agent stands for a biological entity — a molecule, organelle, cell type, organ, behavior.

Prior limitation

Foundation models are frozen encoders — they describe a cell but not what it will do next.

Foundation models become priors; agents provide the trajectory forward in time.

Prior limitation

Prior simulators impose a single global tick; biology spans 15 orders of magnitude in time.

Each agent runs at its natural time scale; Masters synchronize across scales lazily.

Prior limitation

LLM-agent systems hallucinate — Virtual Lab needed a "Scientific Critic" to catch errors.

The Bottom-Up Master is a live simulator verifying every directive against actual lower-level state.

Prior limitation

Every implemented system so far is scoped to one organ or one condition.

The architecture is content-agnostic; the biological entity library is extensible.

Prior limitation

Biological homeostasis is loop-driven, but most systems don't model loops directly.

Feedback loops between the Masters and organism context are the architecture's central invariant.