데브허브 | DEVHUB | BigTool: Agents with large number of tools
Agents dynamically direct their own processes and tool usage. But how many tools can agents reliably use? A large number of tools can introduce cognitive burden, making it challenging for agents to select the right one. BigTool is a prebuilt LangGraph agent that works with large numbers of tools using a simple idea: use a search to retrieve and bind only the most relevant tools to the agent for a given task. We show this working with a simple vectorstore that embeds tool descriptions and uses semantic search to fetch the most relevant tools based upon the user's request. We show this works well even with low capacity local models, like Qwen-14-b.
Video notes:
https://mirror-feeling-d80.notion.sit...
Chapters:
0:00 - Introduction to Tool Calling in Agents
0:47 - Setting Up a Local Agent Demo
1:00 - Running the Agent and Observing Tool Calls
1:27 - The Problem with Traditional Tool Binding
2:00 - The Big Tool Approach
2:27 - How Tool Retrieval Works
3:27 - Benefits of This Approach
3:40 - Big Tool vs. Multi-Agent Architecture
4:58 - Local Model Performance
5:20 - Technical Implementation Details
6:10 - Setting Up the Vector Store
7:00 - Tracing Through the Execution
9:00 - Conclusion and Benefits
![스프링캠프 2024 [Track 2] 1.동시성의 미래 - 코루틴과 버츄얼 스레드 (이상훈)](https://i2.ytimg.com/vi/MnkUX_E9SLg/hqdefault.jpg)
