ls-train

What problem does it solve?

LiteHarness turns an agent or skill into a self-improving system by repeatedly evaluating its behavior, mutating configuration, and keeping or reverting changes based on measurable results.

Core Features & Use Cases

• Autonomous agent training (A/B loop): Spawns two agent variants each cycle, scores their outputs with binary plus qualitative evaluation, and keeps the better configuration while reverting regressions.
• Autonomous skill training (eval-trigger loop): Measures skill trigger accuracy using the Claude Code eval system, then mutates the skill description to improve should-trigger correctness.
• Cycle controls and review gates: Supports pace (fast/medium/deep/adaptive), review intervals for human checkpoints, and a max-cycles stopping condition.

Quick Start

Run the ls-train skill by telling your AI to train an existing target with: “/train --target test-subject --mode agent --review-every 10 --pace adaptive --cycles 0”.