code-act
★ 1.7k·MIT·Updated 2024-05-23
CodeAct unifies LLM agents' actions into an executable code space, enabling dynamic revision and new actions based on execution results. This approach significantly outperforms traditional text and JSON action methods, improving LLM agent success rates on complex tasks.
#LLM Agents#Code Execution#Instruction Tuning#Python#Large Language Models#Coding
01
Features
01Unified Action Space via Executable Code: Consolidates LLM agents' actions into a unified, executable code space.
02Dynamic Action Revision: Allows LLM agents to dynamically revise prior actions or emit new actions based on real-time observations.
03Integrated Python Interpreter: Seamlessly integrates with a Python interpreter for code execution.
04Superior Performance: Outperforms widely used alternatives like Text and JSON in LLM agent success rates (up to 20% higher).
05Instruction-Tuned Agents (CodeActAgent): Provides pre-trained CodeActAgent models (Mistral, Llama-2) that excel in out-of-domain agent tasks.
02
Compatibility
Ollama
Supported
Verified via docs
llama.cpp
Supported
Verified via docs
Kubernetes
Supported
Verified via docs
Docker
Native
Verified via docs
vLLM
Supported
Verified via docs
03
Quick start
1
$ pip install -r requirements.txt
04
Use cases
↳Developing Advanced LLM Agents: For researchers and developers aiming to build more capable and robust LLM agents that can interact dynamically with environments.
↳Automated Code Execution and Problem Solving: For scenarios requiring LLMs to execute code, debug, and iterate on solutions based on execution feedback.
↳Complex Task Automation: For automating multi-turn, complex tasks that benefit from dynamic action revision and tool use.
05
Alternatives
cua★ 30.7k
Open-source infrastructure that gives AI agents computers they can use: sandboxed desktops and VMs, drivers for operating native apps, SDKs and benchmarks for computer-use agents.
rlcard★ 3.6k
Reinforcement Learning / AI Bots in Card (Poker) Games - Blackjack, Leduc, Texas, DouDizhu, Mahjong, UNO.
verl-agent★ 2.4k
verl-agent is an extension of veRL, designed for training LLM/VLM agents via RL. verl-agent is also the official code for paper "Group-in-Group Policy Optimization for LLM Agent Training"
ir-sim★ 1.1k
A Python-based lightweight robot simulator designed for navigation, control, and reinforcement learning
skrl★ 1.1k
Modular Reinforcement Learning (RL) library (implemented in PyTorch, JAX, and NVIDIA Warp) with support for Gymnasium/Gym, NVIDIA Isaac Lab, MuJoCo Playground and other environments
Comments
Log in to leave a comment
No comments yet. Be the first!