Harness² vs Muse Code

Harness²

Refinement needs no grader or answer key, and the code keeps judges out of selection. It runs only through its three benchmark adapters, from a source checkout with no tagged release or CI, and its OpenCode and Codex executors run with unrestricted network access, Codex with its sandbox disabled. The benchmark gains are the authors' claims.

Muse Code

Muse Code runs with approvals and an OS sandbox on by default, per-host network approval, an append-only event log and a headless mode with documented exit codes. It is the only route to Muse Wonder Full. It is closed source with undated release notes and no free route, and subscriptions allow training on code unless Contributor mode is turned off.

Harness²

1. Don't plan to call Harness² as a service. It's a research CLI run from a clone, on JobBench, LAB or WorkBuddy-Bench only
2. Ask the person for a Vertex AI project, model IDs and gcloud auth application-default login before any model-backed run, and start with --max-tasks 1
3. Run executors behind an egress firewall and block huggingface.co and github.com after setup, as the docs advise, so rubrics can't be fetched
4. Write .env yourself with GOOGLE_CLOUD_PROJECT, HARNESS2_SOLVER_MODEL and HARNESS2_JUDGE_MODEL, since .env.example isn't shipped, and never commit it
5. Use the same benchmark, executor, --k, --mode and seed for harness2, harness2-judge and harness2-report, or the results won't line up

Muse Code

1. Set META_API_KEY for headless runs. Browser sign-in isn't available in CI or for Meta Managed Accounts
2. Gate on your own test command, not the exit code. muse exec returns 0 when the turn completes even if tests still fail
3. Pass --max-model-steps on unattended runs so a stuck task can't loop
4. Prefer --disable-approval to --yolo in CI so the sandbox stays on. On Linux the runner needs a working bubblewrap and a non-musl build
5. Turn Contributor mode off in /settings before a subscription session touches private code