Skip to content- Sources: primary, discussion
- Summary: Harness2 runs a Propose-Probe-Compose step that proposes harness edits, probes them on the same task, and composes a refined harness from behavioral contrasts without ground-truth feedback, while a persistent improvement harness carries lessons across tasks. It supports OpenCode and Codex executors and is Apache-2.0. The README warns that executor network access is unrestricted and Codex runs with its sandbox disabled.
- Why it matters: Harness tuning becomes an automated loop you can run, but its default setup executes agents unsandboxed and needs isolation first.