This paper proposes a highly robust autonomous agent framework based on the ReAct paradigm, designed to solve complex tasks through adaptive decision making and multi-agent collaboration. Unlike traditional frameworks that rely on fixed workflows generated by LLM-based planners, this framework dynamically generates next actions during agent execution based on prior trajectories, thereby enhancing its robustness. To address potential termination issues caused by adaptive execution paths, I propose a timely abandonment strategy incorporating a probabilistic penalty mechanism. For multi-agent collaboration, I introduce a memory transfer mechanism that enables shared and dynamically updated memory among agents. The framework's innovative timely abandonment strategy dynamically adjusts the probability of task abandonment via probabilistic penalties, allowing developers to balance conservative and exploratory tendencies in agent execution strategies by tuning hyperparameters. This significantly improves adaptability and task execution efficiency in complex environments. Additionally, agents can be extended through external tool integration, supported by modular design and MCP protocol compatibility, which enables flexible action space expansion. Through explicit division of labor, the multi-agent collaboration mechanism enables agents to focus on specific task components, thereby significantly improving execution efficiency and quality.
The experimental results demonstrate that the autono framework significantly outperforms autogen and langchain in handling tasks of varying complexity, especially in multi-step tasks with possible failures.
| Framework | Version | Model | one-step-task | multi-step-task | multi-step-task-with-possible-failure |
|---|---|---|---|---|---|
autono | 1.0.0 | gpt-4o-mini<br>qwen-plus<br>deepseek-v3 | 96.7%</br>100%</br>100% | 100%</br>96.7%</br>100% | 76.7%</br>93.3%</br>93.3% |
autogen | 0.4.9.2 | gpt-4o-mini<br>qwen-plus<br>deepseek-v3 | 90%</br>90%</br>N/A | 53.3%</br>0%</br>N/A | 3.3%</br>3.3%</br>N/A |
langchain | 0.3.21 | gpt-4o-mini<br>qwen-plus<br>deepseek-v3 | 73.3%</br>73.3%</br>76.7% | 13.3%</br>13.3%</br>13.3% | 10%</br>13.3%</br>6.7% |
one-step-task: Tasks that can be completed with a single tool call.multi-step-task: Tasks that require multiple tool calls to complete, with no possibility of tool failure.multi-step-task-with-possible-failure: Tasks that require multiple tool calls to complete, where tools may fail, requiring the agent to retry and correct errors.The deepseek-v3 model is not supported by
autogen-agentchat==0.4.9.2.
You can reproduce my experiments here.
If you are incorporating the autono framework into your research, please remember to properly cite it to acknowledge its contribution to your work.
Если вы интегрируете фреймворк autono в своё исследование, пожалуйста, не забудьте правильно сослаться на него, указывая его вклад в вашу работу.
もしあなたが研究に autono フレームワークを組み入れているなら、その貢献を認めるために適切に引用することを忘れないでください.
如果您正在將 autono 框架整合到您的研究中，請務必正確引用它，以聲明它對您工作的貢獻.
@article{wu2025autono,
author = {Zihao Wu},
title = {Autono: A ReAct-Based Highly Robust Autonomous Agent Framework},
journal = {arXiv preprint},
year = {2025},
eprint = {2504.04650},
archivePrefix = {arXiv},
primaryClass = {cs.MA},
url = {https://arxiv.org/abs/2504.04650}
}@software{Wu_Autono_2025,
author = {Wu, Zihao},
license = {GPL-3.0},
month = apr,
title = {{Autono}},
url = {https://github.com/vortezwohl/Autono},
version = {1.0.0},
year = {2025}
}pip install -U autonoGet access to unreleased features.
pip install git+https://github.com/vortezwohl/Autono.gitTo start building your own agent, follow the steps listed.
OPENAI_API_KEY# .env
OPENAI_API_KEY=sk-...Agent lets you instantiate an agent.Personality is an enumeration class used for customizing personalities of agents.Personality.PRUDENT makes the agent's behavior more cautious.Personality.INQUISITIVE encourages the agent to be more proactive in trying and exploring.get_openai_model gives you a BaseChatModel as thought engine.@ability(brain: BaseChatModel, cache: bool = True, cache_dir: str = '') is a decorator which lets you declare a function as an Ability.@agentic(agent: Agent) is a decorator which lets you declare a function as an AgenticAbility.from autono import (
Agent,
Personality,
get_openai_model,
ability,
agentic
)@ability
def calculator(expr: str) -> float:
# this function only accepts a single math expression
return simplify(expr)
@ability
def write_file(filename: str, content: str) -> str:
with open(filename, 'w', encoding='utf-8') as f:
f.write(content)
return f'{content} written to {filename}.'You can grant abilities to agents while instantiating them.
model = get_openai_model()
agent = Agent(abilities=[calculator, write_file], brain=model, name='Autono', personality=Personality.INQUISITIVE)agent.grant_ability(calculator)or
agent.grant_abilities([calculator])agent.deprive_ability(calculator)or
agent.deprive_abilities([calculator]) You can change an agent's personality using method change_personality(personality: Personality)
agent.change_personality(Personality.PRUDENT)agent.assign("Here is a sphere with radius of 9.5 cm and pi here is 3.14159, find the area and volume respectively then write the results into a file called 'result.txt'.")response = agent.just_do_it()
print(response)
autonoalso supports multi-agent collaboration scenario, declare a function as agent calling ability with@agentic(agent: Agent), then grant it to an agent. See example.
I provide McpAgent to support tool calls based on the MCP protocol. Below is a brief guide to integrating McpAgent with mcp.stdio_client:
McpAgent allows you to instantiate an agent capable of accessing MCP tools.StdioMcpConfig is an alias for mcp.client.stdio.StdioServerParameters and serves as the MCP server connection configuration.@mcp_session(mcp_config: StdioMcpConfig) allows you to declare a function as an MCP session.sync_call allows you to synchronizedly call a coroutine function.from autono import (
McpAgent,
get_openai_model,
StdioMcpConfig,
mcp_session,
sync_call
)StdioMcpConfig.mcp_config = StdioMcpConfig(
command='python',
args=['./my_stdio_mcp_server.py'],
env=dict(),
cwd='./mcp_servers'
)A function decorated with
@mcp_sessionwill receive an MCP session instance as its first parameter. A function can be decorated with multiple@mcp_sessiondecorators to access sessions for different MCP servers.
@sync_call
@mcp_session(mcp_config)
async def run(session, request: str) -> str:
...@sync_call
@mcp_session('http://localhost:8000/sse')
async def run(session, request: str) -> str:
...@sync_call
@mcp_session('ws://localhost:8000/message')
async def run(session, request: str) -> str:
...McpAgent instance within the MCP session After creating McpAgent, you need to call the fetch_abilities() method to retrieve tool configurations from the MCP server.
@sync_call
@mcp_session(mcp_config)
async def run(session, request: str) -> str:
mcp_agent = await McpAgent(session=session, brain=get_openai_model()).fetch_abilities()
...McpAgent instance and await execution result@sync_call
@mcp_session(mcp_config)
async def run(session, request: str) -> str:
mcp_agent = await McpAgent(session=session, brain=get_openai_model()).fetch_abilities()
result = await mcp_agent.assign(request).just_do_it()
return result.conclusionif __name__ == '__main__':
ret = run(request='What can you do?')
print(ret)I also provide the complete MCP agent test script. See example.
To make the working process of agents observable, I provide two hooks, namely BeforeActionTaken and AfterActionTaken.
They allow you to observe and intervene in the decision-making and execution results of each step of the agent's actions.
You can obtain and modify the agent's decision results for the next action through the BeforeActionTaken hook,
while AfterActionTaken allows you to obtain and modify the execution results of the actions (the tampered execution results will be part of the agent's memory).
To start using hooks, follow the steps listed.
autonofrom autono.brain.hook import BeforeActionTaken, AfterActionTaken
from autono.message import BeforeActionTakenMessage, AfterActionTakenMessagedef before_action_taken(agent: Agent, message: BeforeActionTakenMessage):
print(f'Agent: {agent.name}, Next move: {message}')
return message
def after_action_taken(agent: Agent, message: AfterActionTakenMessage):
print(f'Agent: {agent.name}, Action taken: {message}')
return message
before_action_taken_hook = BeforeActionTaken(before_action_taken)
after_action_taken_hook = AfterActionTaken(after_action_taken)In these two hook functions, you intercepted the message and printed the information in the message.
Afterwards, you returned the message unaltered to the agent.
Of course, you also have the option to modify the information in the message,
thereby achieving intervention in the agent's working process.
agent.assign(...).just_do_it(before_action_taken_hook, after_action_taken_hook)vortezwohl/Autono
October 25, 2024
July 3, 2025
Python