What problem does it solve? Complex multi-step tasks often waste expensive frontier-model calls on trivial steps or fail silently when plans go stale. This Skill runs a meta-orchestration loop where a strong reasoning model acts as conductor while sub-agents run on the cheapest suitable models, with dynamic replanning and budget guards. ## Core Features & Use Cases - RLM Conductor Policy: Assigns planning, judging, and synthesis to the strongest reasoning model while routing worker steps to the cheapest fit per step profile. - Plan-Execute-Verify-Replan Loop: Self-correcting control cycle with convergence criteria, completion guards, and targeted replanning instead of pushing through gaps. - Cost/Quality Governor: Hard budget ceilings per request, task, and run with abort-with-partial-result behavior instead of silent quality degradation. - Recipe Registry: On-demand playbooks for heavy processes including deep research, security audit, competitive analysis, red-team, codebase audit, and benchmarking. - Use Case: Run a full security audit of a codebase where the conductor decomposes the attack surface, cheap long-context workers scan components in parallel, and a strong model judges severity and synthesizes remediation. ## Quick Start Ask the AI to run a deep research or security audit on your target and let the harness plan, delegate to sub-agents, verify results, and replan until convergence.
