How he builds agent systems

Stateless LLM steps, mechanical gate checks, a human editorial pass before anything leaves. Distrusts agents that grade their own work. A claim that matters needs a fetchable primary source or it is dropped.

Provenance: James, in conversation with Claude, Sept–Oct 2026 (unverified). Disposition: stable.

Related: inference-testbed, ai-roundup, research-values

Part of thread: thread-judgement-in-machines