Skip to content

EnvFactory

Hero Section

EnvFactory

Synthetic environments for training robust tool-use agents via trajectory synthesis

Build intelligent agents that master tool use through calibrated synthetic data generation and stateful API verification.

Get Started View on GitHub

Key Features

  • :material-puzzle-outline: Stateful API Environments

    Create realistic, interactive API environments with full state management. Verify agent interactions against specification constraints and catch errors before production deployment.

  • :material-chart-line: Calibrated Trajectory Synthesis

    Generate diverse, high-quality training trajectories with automatic calibration. Balance exploration coverage with task complexity to optimize agent learning efficiency.

  • :material-tools: Tool-Use Agent Training

    Purpose-built tooling for agentic systems that need to master complex multi-step tool interactions. Includes trajectory validation and performance metrics.

  • :material-network: Network-Aware Simulation

    Model real-world dependencies and service interactions. Built on NetworkX for flexible topology definition and graph-based environment composition.

Quick Install

pip install envfactory

What You Can Build

EnvFactory enables you to:

  • Generate synthetic API interactions with calibrated difficulty and state complexity
  • Verify agent behavior against formal specifications before deployment
  • Create diverse training datasets for tool-use agents without manual annotation
  • Simulate realistic failure modes to train robust error-handling strategies
  • Benchmark agent performance across standardized environment scenarios

Getting Started

Ready to train your first agent? Head to our Getting Started guide to:

  • Set up your environment
  • Create your first synthetic API environment
  • Generate training trajectories
  • Train an agent on tool-use tasks

Start Building →

Core Components

envfactory.core : Environment specification and stateful API simulation

envfactory.synthesis : Calibrated trajectory generation and synthesis algorithms

envfactory.agents : Training loops and agent interaction protocols

envfactory.validation : Specification verification and constraint checking

Example Usage

from envfactory import Environment, Agent, TrajectoryConfig

# Create a synthetic API environment
env = Environment.from_spec("payment_api.yaml")

# Configure trajectory synthesis
config = TrajectoryConfig(
    num_trajectories=1000,
    complexity_range=(0.3, 0.9),
    error_injection_rate=0.15
)

# Generate training data
trajectories = env.synthesize_trajectories(config)

# Train your agent
agent = Agent(model="gpt-4")
agent.train(trajectories)

Community

Have questions? Ideas for improvements? Join our community: