Cost-Aware Routing

Matching strategy to task complexity.

The Problem

Tools improve accuracy but add latency. Using the wrong tool wastes time or produces poor results.

Research

"When Do Tools and Planning Help LLMs Think?" (arXiv:2601.02663)

Finding: Tools improve accuracy by +20% but add 40x latency.

Metric	No Tools	With Tools
Accuracy	47.5%	67.5%
Latency	8s	317s

Implications:

1. Simple tasks — Direct response is faster and sufficient
2. Complex tasks — Tools are necessary for accuracy
3. Sweet spot — Match tool complexity to task

arXiv

Tachikoma's Solution

Complexity	Strategy	Latency
Low	Direct response	1-2s
Medium	Single skill	5-15s
High	Multi-skill chain	15-45s
Very High	RLM orchestration	45-120s

Routing Decision

User Request
    ↓
Classify Intent
    ↓
Confidence > 0.7?
    ├── NO → Ask for clarification
    ↓ YES
Context > 2000 tokens?
    ├── YES → Use RLM subagent
    ↓ NO
Task complexity?
    ├── Simple → Direct skill
    ├── Medium → Single skill
    ├── High → Skill chain
    └── Critical → Full orchestration
    ↓
Reflect on approach (freedom to question, flag issues)

Configuration

routes:
  debug:
    confidence_threshold: 0.7
    skill: code-agent
    strategy: direct

  verify:
    confidence_threshold: 0.6
    skill_chain: implement-verify
    strategy: sequential

  complex:
    confidence_threshold: 0.5
    subagent: rlm-optimized
    strategy: rlm

See Also

• Intent Routing
• Research Overview