Agenticoptimizer

Agentic site-occupancy optimizer for atomistic structures. Like a genetic algorithm over lattice sites, but an agent (guided by Skill.md) proposes Na/X swaps instead of GA operators. Structures are relaxed and scored with MatCalc via Main_scripts.py.

Setup

Use an environment with ASE and MatCalc (e.g. the postdoc env):

source /home/mahpe/envs/postdoc/bin/activate
cd /home/mahpe/Code/Agenticoptimizer

For the Cursor agent backend:

pip install cursor-sdk
export CURSOR_API_KEY="cursor_..."

Usage

# Agent-driven optimization (reads config.toml + Skill.md)
python run.py --agent cursor

# Random-swap baseline (no LLM; useful to test the relax loop)
python run.py --agent random --max-iters 3

# Custom config / output
python run.py --config config.toml --out-dir runs/my_run --seed 1 --max-iters 20

CLI options

Flag	Description
--config	Path to config.toml (default: ./config.toml)
--skill	Path to Skill.md
--agent	cursor (Skill.md + Cursor SDK) or random
--model	Cursor model id (default from config: composer-2.5)
--max-iters	Max agent swap steps
--seed	RNG seed for the initial occupancy
--out-dir	Output directory (default: runs/latest)

Config (config.toml)

• cif_path / supercell_size — input structure
• [swap_params] — lattice element, swap species (Na / X), and composition ratio (X = vacancy)
• [relax_params] — MatCalc / ASE relaxation settings
• [agent_params] — iterations, swaps per step, model, skill path, output dir

Outputs

Written under --out-dir (or agent_params.out_dir):

• best.cif — lowest-energy structure found
• step_XXX.cif — structure after each accepted step
• summary.json — energies, occupancies, and swap history
• relaxation.traj — ASE trajectory from the latest relax