AtomisticSkills

Open-source agent harness that empowers general-purpose AI coding agents to conduct atomistic research across materials science, chemistry, and drug discovery through a library of more than 100 curated skills and tools.

   
Affiliation MIT Department of Materials Science and Engineering and Department of Chemical Engineering, with Kookmin University, Harvard, MIT Nuclear Science and Engineering, and Shell
First introduced 2026-05 (arXiv:2605.24002)
Lifecycle stages Multi-stage (literature review, research planning, simulation execution, MLIP benchmarking and fine-tuning, analysis)
Autonomy level Semi-autonomous — general-purpose coding agents orchestrate skills end-to-end under user direction
Domain focus Atomistic research: materials science (53 skills), chemistry (23), drug discovery (18), machine learning (14), general (11)
Availability Open source

Approach

AtomisticSkills hierarchically decomposes scientific workflows for general-purpose coding agents into three layers:

  • MCP tool box — fundamental research tools exposed via the Model Context Protocol (database queries, simulation engines such as FairChem, MatGL, MACE, Atomate2, ORCA, DiffCSP, MatterGen, structure visualization).
  • Skill library — mid-level, semi-flexible atomistic research procedures (e.g., mat-amorphization, mat-stability, mat-diffusion-analysis, ml-foundation-potentials, general-molecular-dynamics) each composed of MCP tools plus declarative SKILL.md specifications.
  • Standards and rules — research, skill, and workflow standards that constrain agent behavior toward reproducible and rigorous practice.

The framework is plug-and-play across coding agents and is designed to be extensible: new skills and MCP servers can be added without modifying the agent.

Validation

The authors validate functional coverage against the scientific literature and demonstrate orchestration across six campaigns:

  • Generative design of Li-ion solid-state electrolytes.
  • High-throughput screening of metal-organic frameworks for CO2 capture.
  • Autonomous MLIP benchmarking and fine-tuning.
  • Multi-stage structure-based virtual screening for drug design.
  • Multimodal X-ray diffraction pattern analysis.
  • Screening of Fe-oxide catalysts for the oxygen evolution reaction.

Notable results

  • More than 100 human-curated multidisciplinary skills spanning database access, thermodynamics and kinetics modeling, MLIP-based and DFT-based simulation engines.
  • Successful end-to-end orchestration of six independent atomistic campaigns by general-purpose coding agents, including an autonomous MLIP fine-tuning loop.
  • Positioned as critical agent infrastructure toward fully autonomous AI scientists in atomistic domains.

Primary paper

Deng et al., “Harnessing AtomisticSkills for Agentic Atomistic Research,” arXiv:2605.24002 (2026).

Other references

None yet.

Code

Described as open source by the authors; repository location to be confirmed.