agent-oraclelisted

# Agent Oracle Five segments covering the design of autonomous agent systems: how you measure them, how you structure them, how you control them, how you launch them, and how you let them improve themselves. ## Every invocation starts with an orchestrator Before any segment work begins, hand the request to an orchestrator first — by name or by segment, this step is not optional and not skippable for a request that looks simple. It is a fixed, bounded decision that terminates in one answer and hands off; it is not a loop, and it does not spawn help to decide. The orchestrator decides four things, in order, each grounded in a specific guide section rather than improvised: 1. **Scope.** Does the request name a segment or clearly belong to one — "design the supervisor topology", "write my launch prompt", "build a regression suite"? Read only that guide and do only that work; do not run the other four, do not walk the person through the sequence, do not append the next stage uninvited. Does it invoke **agent-oracle** generically, or span stages? Work the full sequence below, announcing each segment as you enter it, and do not skip one because it looks satisfied. If it is genuinely unclear which applies, ask once. 2. **Agent count.** One agent, or several? Default to one: multi-agent beats single-agent only when a single agent's context is already degraded, or more compute is deliberately spent (`guides/multi-agent-patterns.md`, "The Token Economics Rea