autoenv-environment-synthesis

name	autoenv-environment-synthesis
title	AutoEnv: Automated Environments for Measuring Cross-Environment Agent Learning
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2511.19304
keywords	["Environment Synthesis","Coding Agents","Automated Testing","Agent Evaluation","DSL-to-Code"]
description	Generate diverse, validated game environments automatically using domain-specific language specifications and LLM coding agents with self-repair, enabling cost-effective (≈$4/env) creation of heterogeneous test domains for evaluating cross-environment agent generalization.

AutoEnv: Automated Environments for Agent Testing

Creating diverse, realistic environments for training and testing agents is expensive and time-consuming. This skill demonstrates how to automate environment generation using a three-layer architectural abstraction combined with LLM coding agents that iteratively synthesize and repair implementation code. The system enables systematic creation of heterogeneous game environments with validated dynamics, observation functions, and rendering pipelines—all driven from high-level YAML specifications.

The core innovation is decomposing environment complexity into three manageable layers (dynamics, observations, rendering) and using coding agents to translate natural language specifications into executable Python code with built-in self-repair loops.

Core Concept

AutoEnv employs a three-layer abstraction architecture to decompose environments systematically:

1. BaseEnv: Implements core dynamics—state space definition, action effects, transition logic, and reward computation
2. ObsEnv: Specializes observation functions through configurable policies that determine what information agents perceive
3. SkinEnv: Applies rendering to convert observations into agent-facing modalities (pixels, text, features)

This layering enables each component to be synthesized and tested independently, with composition handled by the system automatically.

Architecture Overview

• YAML DSL Frontend: High-level environment specification with themes, goals, rules, state variables, and reward conditions
• Prompt Assembly Pipeline: Converts DSL specifications into detailed, contextualized prompts for coding agents
• LLM Coding Agent: Generates implementation code using patterns and examples from libraries
• Three-Stage Verification: Execution testing, level generation validation, and differential model testing
• Self-Repair Loop: Iterative syntax and semantic error correction through LLM-guided debugging

Implementation Steps

The environment synthesis process flows through specification, code generation, validation, and repair stages.

1. Parse and Validate YAML Specification

Convert high-level environment themes into detailed specifications with goals, mechanics, and state variables.

# Example: Simple grid-world specification
environment:
  name: "Treasure Hunt"
   
   
  
      [, ]