Medicine

MARC v1: An Open-Source Multi-Agent Framework for Clinical AI Reasoning and Coordination

Saisha Shetty, Satvik Tripathi, Austin Lin, Colin Zhao, Theodore Kim, Don Enwerem, Jacinta Arnold, Shahriar Faghani, Tessa S Cook

Featured August 15, 2026

Original Paper

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Instead of one big AI trying to do everything, this system uses specialized AI helpers that work together step-by-step, making it easier to see how decisions are made and even letting another AI helper set up their tasks automatically.

In depth

The paper introduces MARC, a multi-agent framework that replaces single, monolithic large language model (LLM) prompts with a structured orchestration of specialized agents for clinical reasoning. This approach enables traceable intermediate outputs and stage-wise failure attribution, enhancing interpretability. A novel Decomposer module automates the generation of task-specific agent prompts from plain language, eliminating manual prompt engineering.

Key Takeaways

1
The framework replaces single LLM prompts with deterministic multi-agent orchestration, allowing specialized agents to handle distinct steps like extraction, reasoning, and answer generation.
2
A Decomposer module automatically generates role-specific prompt templates from plain-language task descriptions, significantly reducing the need for manual prompt engineering.
3
MARC is highly configurable and model-agnostic, supporting both API-based and local CPU-compatible LLM deployments, and enabling non-programmers to modify clinical AI workflows via YAML files.

Conceptual Flow

HIGH LEVEL

Methodology: How Specialized AI Teams Work Together

Instead of one big AI doing everything, this system uses small, specialized AI helpers that pass their work to each other in a clear order.

Big Task Idea

Break Down & Assign

Helper 1's Job

Helper 2's Job

Helper 3's Job

Impact: Easier Setup, Clearer Decisions

This makes it much simpler to set up new tasks without coding, and it's easy to understand why the AI made its final decision.

Hard to Set Up

Confusing Decisions

Make Simple & Clear

Easy Task Setup

Clear Decision Steps

This breakdown was generated by SciGrove. Get the same analysis — intuition, storyboard, peer review, a runnable prototype and a glossary — on any paper you upload or paste a DOI for.