Deep Agents
9 months ago
- #AI Agents
- #Deep Learning
- #Tool Integration
- Using an LLM in a loop to call tools is the simplest agent architecture but can result in 'shallow' agents.
- Deep agents like 'Deep Research', 'Manus', and 'Claude Code' overcome limitations with planning tools, sub agents, file system access, and detailed prompts.
- Deep agents excel in complex tasks over longer time horizons due to better planning and execution capabilities.
- Key characteristics of deep agents include detailed system prompts, planning tools (even no-op ones), sub agents for task division, and file system access for memory and collaboration.
- The 'deepagents' package provides a customizable framework for building deep agents, inspired by Claude Code, with features like system prompts, planning tools, sub agents, and a virtual file system.