Skip to content
TakoAPI
Back to skills

agent-lightning

by olmmlo-cmd

Search & Research
419Downloads reported by clawskills.sh
0Views
0 likesSign in to like this skill
—Not yet rated
Your rating

Brief

Microsoft Research's agent training framework.

Description

What This Skill Does

Agent Lightning is Microsoft Research's framework for training AI agents using Reinforcement Learning, Automatic Prompt Optimization, and Supervised Fine-tuning. It instruments existing agents with minimal code changes via emit helpers or a tracer wrapper. Supports LangChain, AutoGen, CrewAI, and OpenAI Agent SDK out of the box.

Unlike manual prompt engineering or full model retraining, Agent Lightning applies RL and prompt optimization directly to your existing agent code without requiring a rewrite or framework switch.

When to Use It

  • Improving a SQL-generation agent's accuracy through RL training
  • Tuning system prompts automatically for a customer support agent
  • Fine-tuning a coding assistant on preference data
  • Running benchmark evaluations on a multi-agent pipeline
  • Exporting optimized prompts or model weights for production deployment

Example Workflow

INPUT

User asks: train my LangChain agent on SQL generation tasks using reinforcement learning

AGENT
  1. 1

    Instrument the agent by wrapping it with agentlightning tracer or adding emit_tool_call helpers

  2. 2

    Create a config.yaml specifying agent type as langchain, algorithm as grpo, and eval tasks including sql-generation

  3. 3

    Run agent-lightning train --config config.yaml to start training

  4. 4

    Monitor loss curves and reward metrics via agent-lightning dashboard --port 8080

  5. 5

    Export the trained prompts or model with agent-lightning export for deployment

OUTPUT

Optimized agent with improved SQL generation accuracy, exportable for production use

Requirements

  • OPENAI_API_KEY environment variable

Installation

clawhub install olmmlo-cmd/agent-lightning

Run the CLI command, or paste the prompt into Claude Code, Cursor, Codex or whichever coding agent you use.

Links