Agent Oracle is an AI agent orchestration design skill for Claude Code. Before any autonomous work starts, it decides four things and says them out loud: the scope, how many agents the task needs, what stops each loop, and which harness controls and budgets apply.
Each answer is grounded in one of five guides: evaluation, multi-agent patterns, harness engineering, long-horizon prompting and self-improvement loops, backed by cited research and NVIDIA's harness guidance. It defaults to one agent, treats token, tool-call, time and cost limits as a contract, and only trusts a loop bound that is deterministic.
Two standard-library CLIs come with it: an evaluator that scores a JSON test set against a weighted rubric and can gate CI, and a coordination demo. Python 3.10+, no network calls. Part of a three-skill suite with Skill Vision and Research Hound.
Read more: Claude AI skills for research and agent design.