⚡ AgentLightning

Enterprise Agent RL Training & Self-Verification Platform

AgentLightning turns the two most important agent-engineering ideas of August 2026 into a self-hosted platform:

1. Train agents with RL inside the real deployment harness — Microsoft's Agent Lightning v1.0 (Aug 25, 2026) showed that 6,000 samples of harness-based RL lifted Qwen3.5-9B from 41.8% → 56.4% on SWE-bench Verified, with zero changes to business code.
2. Let the model judge itself — LLM-as-a-Verifier topped GitHub trending by demonstrating that best-of-N self-verification lets a 0.11-dollar open-source model beat a closed frontier model.

AgentLightning packages both into a zero-dependency core (pure Python standard library) with a FastAPI control plane, CLI, Python/TypeScript SDKs, a React console, and Docker/K8s/Helm deployment assets.

---

Why AgentLightning?

Problem	AgentLightning's answer
Simulated environments don't transfer to production	Rollout runs the agent in real harnesses (Codex, DeepSeek Harness, Claude Code, local, mock)
Hand-tuned prompts plateau	PPO / GRPO optimize policy parameters against real reward signals
"What is a good answer?" is fuzzy	Pluggable reward functions: verifier (test pass rate), rule-based, LLM-as-a-verifier, safety
A single model is an unreliable judge of itself	Cross-model self-verification: best-of-N + majority voting + verifier-as-judge

Architecture

                ┌─────────────────────────────────────────────┐
                │              Control Plane (FastAPI)         │
                │  /train  /verify  /runs  /harnesses  /healthz │
                └───────────────┬─────────────────────────────┘
        ┌───────────────────────┼────────────────────────┐
        │  CLI (agentlightning) │  Python SDK  │  TS SDK  │  React Console
        └───────────────────────┴──────────────┴──────────┘
                                  │
        ┌─────────────────────────▼─────────────────────────────┐
        │                    Rollout Engine                       │
        │   policy ──► harness (codex/dsh/claude/local/mock) ──►  │
        │              trajectory (steps / observations)          │
        └─────────────────────────┬─────────────────────────────┘
                                  │
        ┌─────────────────────────▼─────────────────────────────┐
        │   Reward (verifier/rule/safety/swebench/composite)     │
        │   Trainer (PPO / GRPO, GAE, LR schedule, early stop)   │
        │   Self-Verify (best-of-N, majority, verifier-as-judge) │
        └─────────────────────────┬─────────────────────────────┘
                                  │
        ┌─────────────────────────▼─────────────────────────────┐
        │  Storage (SQLite) · Observability (Prometheus+W3C)    │
        │  Security (policy/audit/redact) · Sandbox (subprocess/docker) │
        └───────────────────────────────────────────────────────┘

Quick Start

# 1. Clone
git clone https://github.com/huzjie/agentlightning.git
cd agentlightning

# 2. (optional) create a virtualenv and install the API layer
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# 3. Self-check (works fully offline, no keys required)
python -m agentlightning doctor

# 4. Run a training loop in mock mode
python -m agentlightning train --task "Solve: implement a function that sums two integers"

# 5. Self-verify a set of candidate answers
python -m agentlightning verify --task "capital of France" --n 4

# 6. Start the control plane + console
python -m agentlightning serve --port 8000
# open http://localhost:8000/

Configuring Real Providers

Set environment variables (or edit config.json) to use real harnesses and judge models. Double-underscore nesting is supported:

export AGENTLIGHTNING_HARNESS=codex
export AGENTLIGHTNING_PROVIDER=deepseek
export DEEPSEEK_API_KEY=sk-...
export AGENTLIGHTNING_TRAIN__EPOCHS=5
export AGENTLIGHTNING_REWARD__FUNCTIONS='["rule","safety","verifier"]'

Project Structure

agentlightning/          # zero-dependency kernel
  harness/               # codex / deepseek / claude_code / local / mock
  reward/                # verifier / rule / safety / swebench / composite
  policy/                # PolicyModel + checkpoint
  rollout/               # RolloutEngine + sampler
  trainer/               # PPO / GRPO / GAE / scheduler / early-stop
  verify/                # SelfVerifier / best-of-N / majority
  providers/             # 9 LLM providers + mock
  storage/               # SQLite store
  observability/         # Prometheus metrics + W3C trace
  security/              # policy / audit / redact
  sandbox/               # subprocess / docker
  cli/                   # doctor / train / rollout / verify / serve / config
  api/                   # FastAPI control plane (routes/)
  sdk/                   # Python client
sdk/typescript/          # TypeScript SDK
web/                     # React console (zero-build)
k8s/  helm/  Dockerfile  # deployment
docs/                    # guides
tests/                   # pytest suite

Documentation

• Getting Started
• Core Concepts
• Reward Modeling
• Self-Verification
• Training Guide
• API Reference
• Deployment
• Security
• FAQ

License

Apache-2.0. See LICENSE.

---

AgentLightning is an independent project. It is inspired by, but not affiliated with, Microsoft's Agent Lightning or any vendor's Harness products.