Awesome Multi-Agent For Science

A curated list of papers, systems, and resources on multi-agent AI for scientific discovery and research automation.

Table of Contents

• Trend Snapshot
• Scope
• Quality Bar
• By Research Phase
  • Literature Review
  • Hypothesis Formulation
  • Experimentation (Dry Lab)
  • Experimentation (Wet-Lab)
  • Review
• Frameworks And Infrastructure
  • General MAS Frameworks
  • Domain-Specific Platforms
  • Communication And Coordination
• Failure Modes And Open Challenges
• License

Trend Snapshot

OpenAlex-backed trend estimates from 2024 to 2026-04-02.

Period	LLM4Science multi-agent / all	AI4Science multi-agent / all
2024	140 / 1864 = 7.51%	273 / 4804 = 5.68%
2025	415 / 2726 = 15.22%	630 / 6442 = 9.78%
2026 YTD (2026-01-01 to 2026-04-02)	245 / 1145 = 21.40%	322 / 2158 = 14.92%

Methods, keyword families, scripts, and result files are in analysis/trend_of_mas4sci.

Scope

This list focuses on:

• multi-agent LLM systems for scientific discovery
• role-specialized scientific agents and research workflows
• planning, execution, and review in science
• infrastructure, evaluation, and failure modes for scientific MAS

Task tags in the list are:

• phase: Literature Review
• phase: Hypothesis Formulation
• phase: Experimentation (Dry Lab)
• phase: Experimentation (Wet-Lab)
• phase: Review

Scientific discipline is recorded as a domain tag, not a top-level section.

Quality Bar

Every paper entry should:

• be checked against a primary source
• use a real, public, direct URL
• include at least one phase tag and exactly one domain tag
• match the stated Agent pattern
• clearly be about multi-agent for science
• avoid duplicate entries across preprint and venue versions
• include both experimentation tags if a paper genuinely covers both dry-lab and wet-lab execution
• prefer the most specific domain tag over Interdisciplinary

By Research Phase

Verified recent papers, mainly from 2024 to 2026 YTD.

Literature Review

• Two Heads Are Better Than One: A Multi-Agent System Has the Potential to Improve Scientific Idea Generation

  • Agent pattern: collaborative ideation
  • Why it matters: VirSci organizes multiple agents to generate, evaluate, and refine research ideas for autonomous scientific discovery.

• "DIVE" into Hydrogen Storage Materials Discovery with AI Agents

  • Agent pattern: multi-agent extraction and inverse design workflow
  • Why it matters: DIVE organizes agents to read figures and tables from the literature, build a structured database, and support rapid materials discovery.

Hypothesis Formulation

• EvoScientist: Towards Multi-Agent Evolving AI Scientists for End-to-End Scientific Discovery

  • Agent pattern: researcher-engineer-manager
  • Why it matters: EvoScientist combines idea generation, experiment execution, and evolutionary memory in one multi-agent AI scientist workflow.

• Reasoning-Driven Design of Single Atom Catalysts via a Multi-Agent Large Language Model Framework

  • Agent pattern: specialized catalyst-design agents
  • Why it matters: MAESTRO uses specialized LLM roles to iteratively reason about and optimize single-atom catalyst candidates.

• MatPilot: an LLM-enabled AI Materials Scientist under the Framework of Human-Machine Collaboration

  • Agent pattern: human-machine multi-agent collaboration
  • Why it matters: MatPilot uses a multi-agent materials scientist setup to generate hypotheses and experimental schemes.

• MAC-AMP: A Closed-Loop Multi-Agent Collaboration System for Multi-Objective Antimicrobial Peptide Design

  • Agent pattern: closed-loop simulated paper review
  • Why it matters: MAC-AMP uses a multi-agent design-and-review loop to balance activity, toxicity, and novelty in peptide design.

• A Multi-agent Framework for Materials Laws Discovery

  • Agent pattern: multi-agent symbolic regression
  • Why it matters: This framework uses multiple LLM agents to derive interpretable materials laws from scientific data.

• Sci-Mind: Cognitively-Inspired Adversarial Debate for Autonomous Mathematical Modeling

  • Agent pattern: theorist-critic debate with experiential memory
  • Why it matters: Sci-Mind treats mathematical modeling as a debate-and-verification process grounded by retrieval from prior scientific cases.

Experimentation (Dry Lab)

• TianJi:An autonomous AI meteorologist for discovering physical mechanisms in atmospheric science

  • Agent pattern: meta-planner plus worker cohort
  • Why it matters: TianJi decomposes atmospheric science research into planning plus coordinated execution and analysis by specialized agents.

• Multi-Agent Collaboration for Automated Design Exploration on High Performance Computing Systems

  • Agent pattern: job-management, geometry, and inverse-design agents
  • Why it matters: MADA coordinates specialized agents around simulation-driven design exploration on HPC systems.

• Mimosa Framework: Toward Evolving Multi-Agent Systems for Scientific Research

  • Agent pattern: meta-orchestrator plus code-generating agents
  • Why it matters: Mimosa automatically synthesizes and refines task-specific scientific multi-agent workflows with code-generating agents and tool use.

Experimentation (Wet-Lab)

• Validation of an LLM-based Multi-Agent Framework for Protein Engineering in Dry Lab and Wet Lab

  • Agent pattern: conversational multi-agent protein engineering workflow
  • Why it matters: TourSynbio-Agent is validated across both computational and wet-lab protein engineering case studies, making it a strong bridge between dry-lab automation and experimental execution.

• BioMARS: A Multi-Agent Robotic System for Autonomous Biological Experiments

  • Agent pattern: biologist-technician-inspector hierarchy
  • Why it matters: BioMARS integrates LLMs, VLMs, and modular robotics to design, plan, and execute biological experiments with explicit agent specialization.

• AutoLabs: Cognitive Multi-Agent Systems with Self-Correction for Autonomous Chemical Experimentation

  • Agent pattern: self-correcting chemical lab agents
  • Why it matters: AutoLabs turns natural-language goals into hardware-ready liquid-handler protocols and studies why multi-agent chemistry execution works better than simpler baselines.

• Swarms of Large Language Model Agents for Protein Sequence Design with Experimental Validation

  • Agent pattern: decentralized swarm of residue-level agents
  • Why it matters: This work frames protein design as a distributed multi-agent search process and backs it with experimental validation.

Review

• Towards Verifiable and Self-Correcting AI Physicists for Quantum Many-Body Simulations

  • Agent pattern: programming-verifier plus scientific-verifier
  • Why it matters: PhysVEC builds verification and error-correction directly into a multi-agent physics research workflow.

• HLER: Human-in-the-Loop Economic Research via Multi-Agent Pipelines for Empirical Discovery

  • Agent pattern: human-in-the-loop multi-agent pipeline
  • Why it matters: HLER coordinates agents for data auditing, hypothesis generation, econometric analysis, manuscript drafting, and review in one empirical research pipeline.

Frameworks And Infrastructure

General MAS Frameworks

• PiFlow: Principle-aware Scientific Discovery with Multi-Agent Collaboration

  • Agent pattern: principle-aware multi-agent discovery
  • Why it matters: PiFlow frames automated scientific discovery as structured uncertainty reduction guided by principles across multiple scientific domains.

• Open Source Planning & Control System with Language Agents for Autonomous Scientific Discovery

  • Agent pattern: planning-and-control orchestration
  • Why it matters: cmbagent presents a large multi-agent research stack with specialized roles for literature, code, interpretation, and critique.

• OrchMAS: Orchestrated Reasoning with Multi Collaborative Heterogeneous Scientific Expert Structured Agents

  • Agent pattern: orchestrator with heterogeneous expert agents
  • Why it matters: OrchMAS dynamically assembles domain-aware expert agents for long-horizon scientific reasoning tasks.

Domain-Specific Platforms

• AGAPI-Agents: An Open-Access Agentic AI Platform for Accelerated Materials Design on AtomGPT.org
  • Agent pattern: planner-executor-summarizer
  • Why it matters: AGAPI unifies open materials APIs, simulations, and open-source LLMs through a multi-agent orchestration layer.

Communication And Coordination

• MACC: Multi-Agent Collaborative Competition for Scientific Exploration
  • Agent pattern: collaborative competition
  • Why it matters: MACC studies how cooperation and competition can be combined to improve scientific exploration across multiple agents.

Failure Modes And Open Challenges

• Technical failure modes: communication breakdown, coordination overhead, agent disagreement, and cascade failures.
• Scientific reliability issues: hallucination, unverifiable claims, invalid methodology, poor calibration, and reproducibility failures.
• Open problems: long-horizon planning, cross-domain integration, uncertainty handling, efficient coordination, and safety.

License

This repository is licensed under Apache 2.0. See LICENSE.