router_skill

Overview

This skill cost-optimizes AI agent operations by routing tasks to the cheapest model that can handle them. It classifies tasks by complexity and applies tiered model choices, spawning sub-agents when appropriate to save tokens and latency. Use it to reduce spend without sacrificing quality on complex work.

How this skill works

The skill inspects task structure, signals (e.g., debug, summarize), and constraints (tool use, context size, latency) to pick a model tier. It applies escalation rules: downgrade routine work to cheap models, escalate when attempts fail or when high reasoning power or context is required. It can spawn sub-agents with explicit model choices and respects context window and privacy constraints.

When to use it

• Deciding which model to use for a new task or session
• Spawning sub-agents for batched or routine work
• When current model feels like overkill or cost is growing
• Optimizing agent loops where latency and throughput matter
• Enforcing context window and tool-support constraints

Best practices

• Default interactive sessions to a balanced mid-tier model (Tier 2) and downgrade for routine jobs
• Always set an explicit model in sessions_spawn to avoid inheriting an expensive default
• Classify tasks by structure (single-step, multi-step, novel) rather than single keywords
• Batch similar routine tasks in sub-agents to amortize orchestration overhead
• Monitor OpenRouter spend and pin proven cheap model versions to avoid regressions

Example use cases

• Route file reads, formatting, and health checks to Tier 1 or Free models to cut costs
• Spawn dozens of URL-fetch sub-agents on a cheap model for large-scale scraping
• Start debugging on Tier 1, escalate to Tier 3 automatically if previous attempts fail
• Keep agent loops responsive by preferring low-latency Tier 1 models for synchronous steps
• Run nightly background monitoring on Free tier and scheduled reports on Tier 1/2