Harness² vs OpenHands

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.

OpenHands

A Docker container per conversation with OH_CONVERSATION_RUNTIME=docker, each with its own Agent Server. Confirmation mode is off by default in Agent Canvas, and the npm install gives the agent the host's whole filesystem.

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

OpenHands

1. Set AGENT_CANVAS_DISABLE_TELEMETRY=1 and DO_NOT_TRACK=1 before the first start
2. Start Canvas with OH_CONVERSATION_RUNTIME=docker or use the Docker image. The npm install runs the agent on the host
3. Turn on confirmation mode with a risk threshold. Canvas starts with it off
4. Use the SDK or the Agent Server API for headless runs. The openhands --headless CLI is no longer maintained
5. Set filter_tools_regex on the agent to keep unneeded MCP tool definitions out of the context