SPLAT

Simulated Planning with Layered Agent Tasks — a multi-agent simulation framework originally developed as dissertation research, now revamped as a modern Python package.

Agents navigate tile-based environments using a subsumption-style layer architecture: each agent has a priority-ordered stack of behavioral layers, and the highest-priority layer whose conditions are met drives the action for that turn. Environments are authored in Tiled and loaded at runtime; the arcade-based UI provides a live view of the simulation.

Running

poetry install
python -m splat                  # default smoke-test map
python -m splat path/to/map.tmx  # custom map
python -m splat --debug          # verbose agent logging

Controls: R run · P pause · Space step · ↑↓ speed · Q quit

Architecture

• Agents — subsumption layers, affordance-driven actions, team support
• Objects — contextual affordances gated by world state (doors, keys, boxes, …)
• World — turn-based execution, affordance claiming, outcome dispatch
• Tiled loader — converts .tmx maps into simulation state + arcade scene

AI Status

This rewrite from the original Python 2 dissertation code was completed with significant assistance from Claude (Anthropic). Claude models contributed to architecture decisions, code generation, test authoring, and debugging throughout the modernisation effort.