Pi vs Harness²

Pi

Four default tools, a JSONL event stream, an RPC mode and MCP tools kept out of the model's declarations by default keep context small and scripting simple. Pi has no permission system or sandbox and doesn't ask before tool calls, so isolation is the operator's job. Version 1.0.0 is dated 1 October 2026.

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.

Pi

1. Run Pi inside a container or VM for unattended work. It has no sandbox and doesn't ask before running shell commands
2. Use pi --mode json and wait for agent_settled. A failed response doesn't set a nonzero exit code in JSON mode
3. Split the JSONL stream on LF only. Node's readline also splits on U+2028 and U+2029, which are valid inside JSON strings
4. Pass --no-approve or --approve in scripts to make the project trust decision explicit
5. Set PI_TELEMETRY=0 and PI_SKIP_VERSION_CHECK=1, or PI_OFFLINE=1, to stop the default requests to pi.dev

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