Agents

MotleyCrew is first and foremost a multi-agent framework, so the concept of an agent is central to it.

An agent is essentially an actor that can perform tasks. Usually, it contains some kind of loop that interacts with an LLM and performs actions based on the data it receives.

ReAct tool calling agent

MotleyCrew provides a robust general-purpose agent that implements ReAct prompting: motleycrew.agents.langchain.ReActToolCallingMotleyAgent. This agent is probably a good starting point for most tasks.

from motleycrew.agents.langchain import ReActToolCallingMotleyAgent
from langchain_community.tools import DuckDuckGoSearchRun

agent = ReActToolCallingMotleyAgent(tools=[DuckDuckGoSearchRun()])
agent.invoke({"prompt": "Which country currently has more population, China or India?"})

ReActToolCallingMotleyAgent was tested with the newer OpenAI and Anthropic models, and it should work with any model that supports function calling. If you want a similar agent for models without function calling support, look at motleycrew.agents.langchain.LegacyReActMotleyAgent or motleycrew.agents.llama_index.ReActLlamaIndexMotleyAgent.

Using agents from other frameworks

For many tasks, it’s reasonable to use a pre-built agent from some framework, like Langchain, LlamaIndex, CrewAI etc. MotleyCrew provides adapters for these frameworks, which allow you to mix and match different agents together and easily provide them with any tools.

• motleycrew.agents.langchain.LangchainMotleyAgent

• motleycrew.agents.llama_index.LlamaIndexMotleyAgent

• motleycrew.agents.crewai.CrewAIMotleyAgent

Creating your own agent

The simplest way to create your own agent is to subclass motleycrew.agents.parent.MotleyAgentParent.

Note that in a crew, not only an agent can be a worker. A worker is basically any Runnable, and all agents and tools implement the Runnable interface in motleycrew.