Skip to content
weego edited this page Nov 15, 2026 · 7 revisions

LightAgent Wiki

LightAgent is a lightweight Python agent framework for building tool-using agents, memory-enabled agents, MCP-connected agents, Skills-first workflows, deterministic LightFlow workflows, and simple multi-agent collaboration.

The guides below document LightAgent v1.0.0, not later changes on main. Existing English topic pages remain available as companion references.

v1.0.0 Guides / 中文使用指南

Start with installation, then read the safety and deployment guides before going to production. The English and Simplified Chinese guides share identical examples, including prompt text, and document the same version limits.

第一次使用从「安装与第一个 Agent」开始。上线前请阅读安全和部署指南。中英文版使用相同的示例和版本限制。

English 简体中文 Examples / 示例
Installation and Your First Agent 安装与第一个 Agent Native chat, structured results, async / 对话、结构化结果、异步
Tools and Integrations 工具和集成 Python tools, MCP, Skills, storage / Python 工具、MCP、Skills、存储
Memory, Sessions, and RAG 记忆、会话与 RAG Scoped memory, session replay, SQLite retrieval / 记忆隔离、回放、检索
Workflows and Multiple Agents 工作流与多 Agent LightFlow approvals, LightSwarm, offline LightDAG / 审批、交接、离线动态任务图
Agent Backends and Codex 执行后端与 Codex Authenticated local CLI and session resume / 本机 CLI 调用、会话恢复
Durable Execution, Jobs, and Subagents 持久执行、Jobs 与 Subagents Reconnection, budgets, approvals, backups / 重连、预算、审批、备份
Safety, Observability, and Evaluation 安全、观测与评测 Fail-closed Hooks, trace, regression cases, signing / 策略拒绝、追踪、回归、验签
Usage, Cost Estimates, and Budgets 用量统计、费用估算与预算 Token counts, pricing, missing usage / Token 计数、计价、缺失用量
Deployment, Migration, and Verification 部署、升级与验收 CI matrix, wheel checks, LightResearch migration / 环境矩阵、wheel 检查、迁移

Why LightAgent

  • Small Python-native core for agent applications.
  • OpenAI-compatible chat completion API support.
  • Direct Python function tools with OpenAI-style function schemas.
  • Runtime tools, built-in tools, and dynamic tool loading.
  • External memory through a small store() and retrieve() protocol.
  • Optional MCP integration over stdio, SSE, and Streamable HTTP.
  • Skills based on SKILL.md files for reusable task instructions.
  • LightFlow for deterministic DAG-style multi-step workflows.
  • LightSwarm for simple multi-agent routing.
  • Optional structured run results, stream events, and trace events.
  • Dynamic verified workflows with LightDAG, durable tasks with ExecutionService, and an optional Codex CLI agent backend.

Start Here

Goal Page
Install and run the first agent Quick Start
Understand the core runtime Core Concepts
Add Python tools Tools
Add memory safely Memory
Connect MCP tools MCP
Use reusable Skills Skills
Build deterministic multi-step workflows LightFlow
Build multi-agent workflows Multi-Agent
Use Tree of Thought planning Tree of Thought
Enable self-learning behavior Self-Learning
Generate tool code Tool Generator
Use browser automation browser-use Integration
Debug runs and traces Tracing and Debugging
Connect Langfuse Langfuse Observability
Configure providers and local models Model Providers
Check API parameters and return types API Reference
Find runnable examples Examples Cookbook
Upgrade between versions Migration Guide
Understand current constraints Known Limitations
Run tests and CI checks Testing and CI
Deploy an app Deployment Guide
Review production safety guidance Security
Check common answers FAQ
Track planned work Roadmap

Repository Links

Documentation Map

Clone this wiki locally