Harness² vs Qwen 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.

Qwen Code

Apache-2.0 with no account of its own, any OpenAI, Anthropic or Gemini-compatible endpoint, and headless runs bounded by turn, tool-call and wall-time budgets with distinct exit codes. Out of the box, Auto mode lets an LLM classifier approve tool calls, with the sandbox and folder trust both off. Usage statistics are on by default.

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

Qwen Code

1. Pass --approval-mode on every run. The settings default is auto, and one docs page says headless runs default to asking, so don't rely on either
2. Add --max-session-turns, --max-wall-time and --max-tool-calls. All three are unlimited by default. Exit 53 is the turn limit and 55 a budget
3. --yolo doesn't turn on a sandbox. Add --sandbox, or on Linux set tools.executionSandbox with network set to closed
4. Set QWEN_USAGE_STATISTICS_ENABLED=false to stop usage statistics
5. Authenticate with OPENAI_API_KEY, OPENAI_BASE_URL and OPENAI_MODEL in CI. The browser sign-in is discontinued