A high-performance iterative reasoning Agent framework featuring CodeAgent execution pattern and Tool Search Tool (TST) for dynamic tool discovery and efficient arbitrary query processing.
Minion is a Python Agent framework centered on iterative reasoning, self-described as "Agent's Brain," currently in Alpha stage (v0.1.21). Its core design revolves around the Brain reasoning loop (Query → Solution → Check → Improve), offering two primary usage paradigms: the recommended CodeAgent (Python code execution + tool calling) and direct Brain.step() invocation.
For context and token optimization, Minion implements three key mechanisms: Tool Search Tool (TST) for dynamic tool discovery to reduce token consumption, Auto-compact for automatic context window management via history summarization, and Auto-decay for auto-saving oversized tool responses after TTL expiration. The framework supports multiple backends including OpenAI, Azure OpenAI, Anthropic Claude, AWS Bedrock, and LiteLLM (100+ providers), unified through a providers abstraction layer. Extensibility features include a modular Skills system, MCP protocol tool integration, optional browser automation via browser-use, and a Gradio Web UI. The configuration system supports project/user-level config layering with environment variable injection (${VAR_NAME} syntax, env_file, environment inline). Deployment options cover PyPI installation (pip install minionx), source installation, and Docker.
Use cases include mathematical reasoning, code generation and execution, competition-level reasoning (e.g., AIME), general Agent task orchestration, MCP tool orchestration, multi-Agent collaboration (ensemble workflows like multi-worker + majority voting), and browser automation tasks.
On benchmarks, the project claims GSM8K at 96% (DeepSeek-Chat), HumanEval at 98.2% pass@1 (GPT-4o), Game of 24 hardest 20 problems at 100%, and AIME 2024 at 26% (4/15), though none have been independently verified by third parties. The TST "85% token reduction" claim also lacks independent benchmark evidence.
Installation
pip install minionx
pip install minionx[all] # all optional dependencies
Quick Start — CodeAgent
from minion.agents.code_agent import CodeAgent
agent = await CodeAgent.create(
name="Minion Code Assistant",
llm="your-model",
tools=all_tools,
)
async for event in await agent.run_async("your task here"):
print(event)
Quick Start — Brain
from minion.main.brain import Brain
brain = Brain()
obs, score, *_ = await brain.step(query="what's the solution 234*568")
print(obs)
Caveats: Alpha stage with no API stability guarantees; no standalone website, documentation lives entirely in the GitHub repo docs/ directory; Python compatibility is 3.8–3.11, 3.12+ not confirmed; multimodal support scope is unclear; related projects minion-agent and minion-code are mentioned in README but no specific repo URLs provided.