Agentic Science Worker

An autonomous AI agent system for computational materials science research. Designed to work with multiple coding agents (Claude Code, Aider, OpenAI Codex, Cursor), this system enables AI to conduct scientific research like a PhD-level computational scientist.

Overview

The Agentic Science Worker can:

• Run molecular dynamics simulations (LAMMPS) with literature-sourced parameters
• Perform DFT calculations (Quantum ESPRESSO) for electronic structure
• Search scientific literature and extract methodology/parameters
• Query materials databases (Materials Project) for structures and properties
• Analyze results and compare with published values
• Execute on HPC clusters for large-scale computations

The agent operates autonomously: given a scientific question, it researches the methodology, finds parameters, runs simulations, verifies results against literature, and iterates until achieving physically reasonable results.

Architecture

┌─────────────────────────────────────────────────────────────┐
│                     Coding Agent                            │
│    Claude Code │ Aider │ OpenAI Codex │ Cursor              │
│     (AGENTS.md defines researcher behavior and methodology) │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                        Skills                               │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐       │
│  │ LAMMPS   │ │ QE       │ │ HPC      │ │ MLIP     │       │
│  │ Skill    │ │ Skill    │ │ Skill    │ │ Skill    │       │
│  └──────────┘ └──────────┘ └──────────┘ └──────────┘       │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐                    │
│  │Literature│ │Materials │ │ Data     │                    │
│  │ Search   │ │ Database │ │ Analysis │                    │
│  └──────────┘ └──────────┘ └──────────┘                    │
└─────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                    External Tools                           │
│  LAMMPS │ Quantum ESPRESSO │ Python │ HPC Cluster │ Web    │
└─────────────────────────────────────────────────────────────┘

Supported Agents

All agents read AGENTS.md (the industry standard) as their primary context file.

Quick Start

Prerequisites

• A supported coding agent:
  • Claude Code CLI with subscription, OR
  • Aider with API key, OR
  • Cursor
• Python 3.10+
• LAMMPS (with GPU support recommended)
• Quantum ESPRESSO (optional, for DFT)
• Materials Project API key

Installation

1. Clone the repository:

git clone https://github.com/yourusername/agentic-science-worker.git
cd agentic-science-worker

1. Copy and configure settings:

cp config.example.yaml config.yaml
cp .claude/settings.json.example .claude/settings.json
cp .mcp.json.example .mcp.json

1. Edit configuration files with your paths and API keys:

# config.yaml
binaries:
  lammps: "/path/to/your/lammps/bin/lmp"
  qe_cpu: "/path/to/qe/bin"

api_keys:
  materials_project: "YOUR_MP_API_KEY"

1. Verify infrastructure:

cd benchmarks/evaluation
python harness.py --verify

Running the Agent

With Claude Code:

cd /path/to/agentic-science-worker
claude

With Aider:

cd /path/to/agentic-science-worker
aider --read AGENTS.md

With Cursor:

cursor .
# Uses AGENTS.md and .cursorrules automatically

Example prompts (any agent):

Calculate the self-diffusion coefficient of liquid argon at 94K
Find the lattice constant of copper using the Mishin EAM potential
Calculate the band structure of silicon

Benchmark Suite

The project includes a comprehensive benchmark suite to evaluate agent capabilities:

Tiers

...