fhaviary

What it is and what it does

This package provides a gymnasium-style framework for building custom reinforcement learning environments where language model agents solve tasks by making tool calls and exchanging messages. It defines core abstractions—Environment, Message, Tool, and ToolRequestMessage—that handle agent-environment communication following OpenAI's message standard. The library ships with pre-built environments for math, general knowledge, biological sequences, scientific literature, and protein stability tasks, plus a Jupyter notebook environment.

The core workflow involves subclassing Environment to define custom tasks, implementing reset() to initialize state and tools, and step() to execute agent actions and return observations and rewards. Communication flows through a message formalism where agents emit ToolRequestMessage objects containing tool calls, and the environment responds with ToolResponseMessage objects or general observations. The framework is designed to integrate with a sister library for defining agents as Language Decision Processes, though it remains task-agnostic and works with any agent implementation.

Use it for

• Train language agents on math reasoning tasks by defining environments with calculator tools and evaluating agent trajectories
• Build custom multi-step reasoning environments where agents call domain-specific tools to solve problems
• Evaluate language model performance on knowledge retrieval by wrapping datasets as environments with tool-based interaction
• Research agent behavior on scientific tasks like protein stability prediction or sequence analysis
• Run Jupyter notebooks as interactive environments where agents execute code and observe outputs

Worth the install?

AI-flagged interpretation of the facts on this page. Verify before relying on it.

Install

fhaviary on PyPI

Before you install

Low install friction with a pure-Python wheel distribution. Actively maintained with recent commits. Supports Python 3.11 through 3.14 with 5 runtime dependencies including httpx and pydantic.

Requires Python 3.11 or later. A custom agent implementation or integration with an external agent library is needed to run agents on the environment; this package defines the environment interface only.

License in practice

Apache License 2.0 is permissive and allows commercial use, modification, and distribution with minimal restrictions. You must include a copy of the license and note any changes to the source.

Quickstart

pip install fhaviary

from aviary.core import Environment, Message, ToolRequestMessage, Tool

class CounterEnv(Environment):
    async def reset(self):
        self.state = {"count": 0}
        self.tools = [Tool.from_function(self.increment)]
        return [Message(content="Start counting")], self.tools
    
    async def step(self, action: ToolRequestMessage):
        obs = await self.exec_tool_calls(action)
        return obs, 0, False, False
    
    def increment(self):
        self.state["count"] += 1
        return f"count={self.state['count']}"

Package facts

Evidence: fhaviary-0.35.0-py3-none-any.whl

Tags

More Artificial Intelligence packages

See also gymnasium · nemo-gym · gem-llm · gym-aloha · pettingzoo · livekit-plugins-anthropic · stable-baselines3 · kaggle-environments · openenv-core · TextArena

Further reading