The layer that makes an agent safe to leave running: an explicit orchestration graph, evaluation that gates deploys, capability scopes enforced at call time, and cost you can attribute per task.
Orchestration design
Task decomposition, tool contracts, retrieval and memory, and a graph that a long-running agent can be resumed from rather than restarted.
Evaluation harnesses
Golden sets, scoring rubrics and regression gates in CI, so a prompt change cannot quietly degrade a path nobody read that week.
Capability scoping
Tools behind grants checked at call time instead of asked for in a prompt, with human confirmation on anything irreversible.
Model routing & cost
The cheapest adequate model per hop, hard budget ceilings, and cost attributed per task rather than per invoice.