Knowledge Vault

Tiered AI Routing

3 min read

Overview

Tiered AI task routing is a defense-in-depth strategy for AI workload management that directs tasks to the specific model best suited for each job based on task-model fit — complexity, trust level, context size, and cost tolerance. The fundamental insight is that not all AI tasks require the same computational power: routing everything to the most capable model wastes expensive compute on trivial work while failing to protect critical assets from lower-trust models. The routing mechanism uses labels as strategic gates that enforce cost optimization, security boundaries, and quality thresholds simultaneously.

Key Concepts

  • Five-tier architecture — Tier 1: Critical reasoning and security (Claude Code Max — cc-arch, cc-security). Tier 2: High-context research (Gemini CLI / Google Deep Research — gem-docs, gem-research, gem-jules). Tier 3: Bulk execution (Aider, Copilot CLI, OpenCode — aid-refactor, cop-bg, oc-bulk). Tier 4: Low-cost exploration (Qwen, Gemini free — qc-spike, gem-quick). Special tier: friction label routes to human knowledge collection, not AI.
  • Labels as routing directives — Labels are not metadata tags; they are strategic gates enforcing task classification, model selection, cost allocation, and trust boundary. The cc-security label is a hard mandatory boundary for authentication, authorization, and cryptography — cannot be overridden for cost reasons.
  • cop-bg label — Routes to GitHub Copilot CLI (Sonnet). Use pattern: agent reads .management/copilot-tasks.md, claims a task via branch creation, works autonomously. This is the mechanism that powers the multi-agent workflow pattern.
  • Task-model fit matrix — Tier 1 handles architectural complexity with moderate context; Tier 2 handles comprehensive research with massive (1M+) context; Tier 3 handles repetitive multi-file changes; Tier 4 handles exploratory throwaway work.
  • Anti-pattern: undifferentiated routing — Sending everything to the most capable model wastes compute, accumulates unnecessary cost, violates security boundaries, and starves critical tasks of premium resources when they need them.
  • friction escape hatch — Routes to Peppermint (human knowledge log), not AI. Flags “Silent Struggle” or broken processes for capture. Enforces the principle: never make the same mistake twice.
  • Batch by tier — Context switching between tiers carries overhead (different API auth, prompt formats, output quality expectations). Batch tasks by tier to minimize switching cost.
  • Multi-Agent Workflow — The cop-bg tier is the operational layer this routing framework coordinates
  • AP Ecosystem — The ecosystem this routing strategy governs
  • Agent Patterns — Foundational agent architecture concepts underlying tier design
  • Ops-Core — Governance layer that enforces the security boundaries (cc-security, cc-arch) at the top of the routing hierarchy

Graph View

Backlinks