A Rust agent harness: run an AI agent from a typed task contract to a verified result. Provider-agnostic (OpenRouter, Anthropic, OpenAI) with fallback between them, multi-file edits with grep/find over a workspace, budgets, classified provider failures with kind-aware retry and backoff, stall detection with a bounded replan, full trace, resumable runs, execution-based verification, a layered permission policy with a human-approval gate, contained sub-agent composition, an OS-native/OS-neutral sandbox (macOS sandbox-exec, Linux namespaces, portable floor; Windows wall-clock only) isolating model-produced code per run, durable checkpoint/resume for unattended runs, an MCP client (stdio and streamable HTTP), a deny-by-default network egress policy, budget-aware context assembly that compacts superseded observations and re-reads what a later write invalidated, and durable cross-run memory keyed to the workspace. Embeddable in-process.