Text worlds for interactive agents

WordPlay

A modular toolkit for building text-based simulation environments where agents move, act, communicate, and cooperate through text and tile-based worlds.

↗ Code

Full Melting Pot suite 4 composable primitives Text and tile rendering

WordPlay RPG town demo with in-world chat and trade — RPG Town

~ ~ ~ ~

Collaborative Cooking Figure Eight — Cooking: Figure Eight

Predator Prey Orchard — Predator-Prey: Orchard

Chemistry Three Metabolic Cycles — Chemistry: 3 Cycles

Commons Harvest Partnership — Harvest: Partnership

Paintball Capture the Flag — Paintball: CTF

Externality Mushrooms Dense — Mushrooms: Dense

Allelopathic Harvest Open — Allelopathic Harvest

// Abstract

Agents do not live in the real world, so WordPlay gives them worlds made of words. WordPlay is a Python framework for constructing environments for interactive agents. Environments are built from four composable primitives: Entities, Actions, Systems, and Movement. This makes it easy to go from a tiny grid world to a bustling multi-agent environment with inventory, combat, communication, and trading. Environments can be rendered as text descriptions or tilemaps, and agents can be controlled by language models, scripted policies, or humans. The framework also includes text-grid adaptations of DeepMind's full Melting Pot scenario suite.

~ ~ ~ ~

// How It Works

Under the hood, every environment follows the same loop: observe, choose an action, validate it, execute it, run shared systems, and render the next state.

WordPlay environment loop: agents observe, choose actions, validate them, update modular systems, and render the next state

The same interface supports human policies, scripted policies, and LLM policies, so you can swap the driver without rebuilding the world. Components and systems can be mixed, matched, and extended without pulling the rest apart.

~ ~ ~ ~

// What makes it different

⚙

Composable primitives

Entities, actions, systems, and movement models snap together like building blocks. Swap the movement system without touching the combat code.

🤖

LLM-native design

Agents act from natural-language action descriptions. Plug in OpenAI, OpenRouter, or local models through one shared interface.

👁

Human-in-the-loop

Drop in a human policy for interactive play, debugging, or data collection. Same world, different driver.

📄

Text & tile rendering

Render environments as text descriptions or as 2D tilemaps with sprite assets. Replay and live-view rendering are built in.

🤝

Multi-agent ready

Built for multi-agent settings with communication, trading, shared tasks, and the occasional glorious mess of coordination.

📈

Built-in benchmarks

Text-grid adaptations of DeepMind's full Melting Pot scenario suite — some examples below.

~ ~ ~ ~

// Examples

A few small worlds to get you moving: chat, trade, and a single-point room where three agents share one abstract location.

Single-point summit with agents ringed around a round table taking turns speaking

~ ~ ~ ~

// Benchmarks

WordPlay includes text-grid adaptations of DeepMind's Melting Pot — the full scenario suite, reimagined as text and tile-based worlds.

Boat Race Eight Races — Boat Race: Eight Races

Chemistry Three Metabolic Cycles With Plentiful Distractors — Chemistry: Three Cycles + Distractors

Chemistry Two Metabolic Cycles — Chemistry: Two Metabolic Cycles

Chemistry Two Metabolic Cycles With Distractors — Chemistry: Two Cycles + Distractors

Collaborative Cooking Asymmetric — Collaborative Cooking: Asymmetric

Collaborative Cooking Circuit — Collaborative Cooking: Circuit

Collaborative Cooking Cramped — Collaborative Cooking: Cramped

Collaborative Cooking Crowded — Collaborative Cooking: Crowded

Collaborative Cooking Forced — Collaborative Cooking: Forced

Collaborative Cooking Ring — Collaborative Cooking: Ring

Commons Harvest Closed — Commons Harvest: Closed

Commons Harvest Open — Commons Harvest: Open

Factory Commons Either Or — Factory Commons: Either-Or

Fruit Market Concentric Rivers — Fruit Market: Concentric Rivers

Paintball King Of The Hill — Paintball: King of the Hill

Predator Prey Alley Hunt — Predator-Prey: Alley Hunt

Predator Prey Open — Predator-Prey: Open

Predator Prey Random Forest — Predator-Prey: Random Forest

~ ~ ~ ~

// Getting Started

Two commands, one grid world, one human-controlled agent:

                # install the base package (renderer and model extras are optional)
                $ pip install -e .

                # run a simple environment with a human-controlled agent
                $ python examples/simple_env_0.py

See the GitHub repo for the current code, examples, and setup details.

~ ~ ~ ~

// Citation

If WordPlay is useful in your work, reference the project repository:

@software{wordplay_2025, author = {WordPlay contributors}, title = {WordPlay: A Modular Framework for Text-Based Environments}, year = {2025}, publisher = {GitHub}, url = {https://github.com/AndreiMuresanu/Word_Play} }

// Thanks

The name "WordPlay" was suggested by Darci Prout.

Sprite assets in the tilemap renderer come from open-licensed packs including DawnLike, Kenney Tiny Dungeon, and Mana Seed Character Base. See sprite_library/art_source/citations.md for full credits.