Iterative agent harness improvement: run a coding agent on a hard task, generate the reusable tooling it was missing, qualify it, and replay fresh sessions with it activated. Works with Codex and Claude Code.
Skill Codex
Optimize HarnessGym tensor-layout kernelplan.json tasks with a generated MCP server for plan validation, dev/final benchmarking, trace analysis, rollback-safe search, candidate application, history comparison, and experiment ranking. Use when a task asks to minimize bestcycles for benchmark.py or tune tensor…
Skill Claude CodeCodex needs its repo
Use for the H100 Triton fused RMSNorm + SiLU gate optimization task. Provides the workflow and MCP tooling for objective runs, rollback-safe config sweeps, source diagnostics, benchmark history, and final held-out verification.