Every agent action is an inference call.

For a few years, inference meant a person typing a prompt and reading the answer. That's changing. A coding agent working through a task makes many calls; an agent framework runs many agents; an organization built from agents runs around the clock with nobody watching each request.

The biggest change isn't the volume. It's who reads the error. A person who sees "service unavailable" waits and tries again later. An agent does what its SDK tells it, and most SDKs retry. When an API returns a retryable error for a problem retrying can't fix, an agent loops through its budget. Infrastructure for that caller has to be judged on whether it fails safely, not just on price.

Gatewayz sits below the agent stack, not inside any one of it. That's what makes neutrality credible: we don't build agents and we don't sell a model, so we have no reason to favor one runtime or one provider over another.

• Neutral by construction. The model you ask for, or a clear refusal. Switching models is a one-word change; switching away from Gatewayz is a one-line change.
• Machine-legible. Every failure tells software whether to retry, stop or escalate.
• Measured, not claimed. Numbers go on this site only after they've been measured, with the method published next to them.

• Published benchmarks against calling providers directly, once production runs exist.
• Same-model failover proven on live traffic, with the evidence published.
• A second side of supply: independent GPU operators serving verified work alongside the major providers — see compute providers.