# microsoft/agent-lightning

The absolute trainer to light up AI agents.

Repository: https://github.com/microsoft/agent-lightning
Canonical: https://ross.abutalabs.com/products/agent-lightning
Homepage: https://microsoft.github.io/agent-lightning/
Language: Python
License: MIT
License Family: permissive
Topics: agent, agentic-ai, llm, mlops, reinforcement-learning
Last push: 2026-08-26T09:50:07+00:00

## Health v2 (maintenance only)
Score: 85/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 99, longevity 31
- inputs: {"age_days": 441, "days_push": 7, "days_rel": 9, "gap_med": 10.0, "n_releases_24m": 9}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 17863, forks 1570 (observed 2026-08-28T04:11:20.485101+00:00)

## What it is
Agent Lightning is a lightweight (~3,500-line) agentic reinforcement learning framework from Microsoft for training LLM-powered agents with their real harnesses. It proxies model requests to capture training data, runs rollouts locally or as Kubernetes Jobs, and uses verl/vLLM for policy updates.

## Use cases
- train an LLM agent with reinforcement learning
- fine-tune a coding agent on SWE-bench tasks
- run RL training rollouts on Kubernetes
- capture agent interactions as training data without changing agent code
- improve a Qwen model's agentic coding performance with RL

## When to choose
- you want to RL-train agents while keeping their existing tools, control flow, and environments unchanged
- you need Kubernetes-native rollout execution without external sandbox services
- you want a minimal, hackable RL training stack built on verl and vLLM

## When to avoid
- you only need simple supervised fine-tuning or prompt engineering rather than agentic RL
- you lack GPU infrastructure or don't want a verl/vLLM dependency stack
- you need a turnkey commercial product rather than a research-oriented framework

## Facets
- artifact type: framework
- maturity: active
- function: llm-training, reinforcement-learning, agent-framework, llm-inference
- domain: large-language-models, reinforcement-learning, machine-learning
- platform: python
- tags: agentic-rl, verl, vllm, mlops, rl-training, agent-harness, swe-bench, ai-agents, kubernetes, docker, gpu, linux

## Member repositories
- microsoft/agent-lightning (main) score 85

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:11:20.485101+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:02:26.043328+00:00, confidence not recorded.
  - readme: https://github.com/microsoft/agent-lightning (fetched 2026-08-28T04:11:20.485101+00:00, sha a0fd16ebb7cc)
  - homepage: https://microsoft.github.io/agent-lightning/ (fetched 2026-08-29T08:00:25.881876+00:00, sha 0cc59130a56d)
- Data as of 2026-08-30T08:39:29.467469+00:00.