Agentic race

Run competing models on one task and have a judge select the strongest answer.

Run multiple configured models concurrently:

arka race "design a test plan for this API" \
  --models openrouter/model-a,ollama/qwen3:8b,groq/llama-3.3-70b-versatile \
  --judge openrouter/model-a --json

Each contestant is bound to its declared provider/model. The judge receives only the task and contestant outputs and scores correctness, completeness, testability, and safety. Failed contestants are retained in the report, and a judge failure is explicit rather than silently selecting an answer.

Natural language works: “race several agents on a coding task.” Use the result as a recommendation; Arka does not write, deploy, register, or submit anything automatically from a race winner.