Participate

Iterate locally with the ha CLI, then submit for the scored run.

Install the haCLI, clone the template repo, and tune your harness against your own key — instant feedback, no waiting. When it's ready, ha submit hands it to the server, which judges it and queues the official, org-funded eval.

# 1. install the ha CLI
$ curl -fsSL https://get.harnessarena.xyz | sh

# 2. authenticate (device-code GitHub login)
$ ha login

# 3. clone the template repo
$ ha clone                      # AGENTS.md / skills/ / agents/

# 4. iterate locally on your own key + harbor
$ ha run --local --tasks fix-git  # needs OPENROUTER_API_KEY

# 5. submit — the server judges + queues the scored run
$ ha submit

# 6. watch the leaderboard

CLI iteration guide →How it works

What gets scored

Your harness — the plugins, skills, agent profiles, and config you layer over vanilla pi — is run across all 89 Terminal-Bench tasks on org-funded infrastructure. Your score is the billed OpenRouter dollars from the generation ledger, and you rank among everyone who matches or beats the GLM 5.2 baseline pass count. Lowest qualifying spend wins.

No task-specific content is allowed: the LLM judge and the tripwire scan enforce harness-only engineering. See the full pipeline for how ha submit is screened, queued, and scored on org-funded infra.