A Rust agent harness: run an AI agent from a typed task contract to a verified result. Provider-agnostic (OpenRouter, Anthropic, OpenAI), multi-file edits with grep/find over a workspace, budgets, retry, full trace, resumable runs, execution-based verification, a layered permission policy with a human-approval gate, contained sub-agent composition, and an OS-native/OS-neutral execution sandbox (macOS sandbox-exec, Linux namespaces, Windows Job Objects, portable floor) that isolates model-produced code per run. Embeddable in-process.