One gateway, many models — and a round-robin where agents grade each other

The idea

Review is better with more than two or three model families in the pool. The wrong way to get there is wiring providers one at a time. The right way is to wire one AI gateway and get many models at once: Grok, DeepSeek, Mistral, Llama, Qwen and others, behind one key and one code path.

Gateway options compared

• A hosting platform's AI gateway — unified, observable, with model fallbacks and zero retention.
• An open model router — the broadest catalogue, OpenAI-compatible, hundreds of models behind one key.
• One vendor direct — a single model family, the worst return. Rejected.

Building it the disciplined way

1. The gateway key goes into the vault through the sealed paste flow. It's read server-side into memory only.
2. A gateway branch on the agent-run tool. The model name passes straight through.
3. Usage is metered like every other provider, with no silent provider swap.
4. Deploy, then verify with a live round-trip.

Round-robin peer review

With a gateway, any N models are available. Agents A, B and C generate independently under anonymous labels, and agent D grades. Which model grades rotates on each review, so the models grade each other. Provider names are hidden from the grader so the brand can't bias the score.