Antigravity CLI vs Harness²

Antigravity CLI

Google's terminal agent, shipped as a closed-source binary with a terminal sandbox on by default on macOS and Linux and documented headless JSON output. Interactions are used to improve Google products unless users opt out, and a headless run exits 0 when a tool is denied.

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.

Antigravity CLI

1. Read the status field from --output-format json and read stderr too. A tool refused for lack of approval still exits 0
2. Pre-approve each shell command and tool a headless run needs with action(target) rules under permissions.allow in ~/.gemini/antigravity-cli/settings.json, because headless runs cannot ask
3. Export GEMINI_API_KEY after setting modelProvider to gemini. The CLI won't start when the key is missing, and setting the variable alone has no effect
4. Turn off Enable Telemetry in the Account settings before running on private repositories if Interactions should not be used for product improvement
5. Set --print-timeout explicitly. The docs give a five-minute default, while agy --help shows 0s

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