Microsoft has shipped the first official release of the Agent Lightning Skill (v1.0.1), a drop-in tool that lets a coding agent automatically optimize other AI agents. Released on August 24, it installs directly into Claude Code, Codex, and GitHub Copilot and tunes an agent's prompts, tools, workflows, model choice, and reasoning settings while balancing accuracy, cost, latency, and reliability.
Try It: Optimize an Agent in Your Existing Coding Tool
Install the skill with one command in your terminal: gh skill install microsoft/agent-lightning agent-lightning --agent claude-code (swap claude-code for codex or copilot). Once installed, point it at an agent you are building and describe the metric you care about, then the skill walks your coding agent through systematic changes to prompts, tool definitions, and model settings. The underlying Agent Lightning framework is a roughly 3,500-line reinforcement learning system that trains agents through their real harness with no code changes, so the skill applies that optimization loop without you writing any training code.
Why It Matters for Creators
Most people building agents tune them by hand: rewrite the prompt, swap a tool, try a cheaper model, eyeball the result, repeat. Agent Lightning turns that guesswork into a guided loop your existing agent runs for you. In Microsoft's own benchmark, the approach lifted a Qwen3.5-9B coding agent from 41.8 to 56.4 percent on SWE-bench Verified using only 6,000 training samples, evidence that structured optimization beats manual trial and error. For anyone shipping an agent into production, that means fewer wasted iterations and a clearer path to a cheaper, faster, more reliable build.
Key Details
Release: Agent Lightning Skill v1.0.1, published August 24, 2026.
Supported agents: Claude Code, Codex, and GitHub Copilot.
What it tunes: prompts, tools, workflows, model choice, and reasoning settings.
License: MIT (open source). Full method in the Agent Lightning research paper.
What to Do Next
Read the official documentation to see the supported metrics and harness setup, then run the skill against a small agent before wiring it into a larger pipeline. If you are already coordinating coding agents inside your team chat, see our coverage of GitHub Copilot's agentic sessions in Slack and Teams, or compare it against the open-source TrueForge agent harness.
