## Problem
When Agents execute complex tasks, all steps are typically handled by the same large language model (LLM). However, many steps are actually simple—such as formatting output or extracting parameters—and can be fully accomplished by much cheaper smaller models.
How can costs be reduced without sacrificing quality?
## RSI-Router’s Solution
RSI-Router (Evolving Subtask-Level LLM Routing) proposes intelligent routing at the subtask level, dynamically selecting the most suitable model for each step.
### Core Idea
A complex task can be decomposed into multiple subtasks, each with varying difficulty:
```
User request: "Analyze this code and generate documentation"
Subtask decomposition:
1. Read code → Simple (information extraction) → Small model
2. Understand logic → Complex (requires reasoning) → Large model
3. Generate documentation → Simple (formatted output) → Small model
4. Check consistency → Complex (requires understanding) → Large model
```
### Three Key Technologies
**1. Subtask-Level Routing**
Instead of routing the entire task to a single model, each substep is decided independently. The system analyzes the nature of each step:
- Information extraction, formatting → Small model
- Reasoning, analysis, creation → Large model
**2. Evolving Skill Learning**
The routing strategy is not static. The system learns from historical executions:
- Which types of steps suit which models
- Dynamically adjusts routing thresholds
- Continuously optimizes with usage
**3. Cost-Aware Optimization**
Minimizes cost while guaranteeing quality:
- Builds a cost-quality trade-off model
- Automatically identifies the optimal balance point
- Supports configurable quality lower bounds and cost upper bounds
### Experimental Results
- Cost reduction of 60–80%
- Performance degradation less than 2%
- Latency reduction of 30–50%
### Real-World Application Scenarios
**Customer Service Agent:**
- Understanding user queries → Large model
- Querying knowledge base → Small model
- Generating responses → Model selection based on complexity
- Logging conversations → Small model
**Code Assistant:**
- Understanding requirements → Large model
- Code completion → Large model
- Formatting output → Small model
- Generating documentation → Small model
### Implications for Holon
Holon can implement a similar routing mechanism:
- Identify simple vs. complex steps within a task
- Automatically assign them to different models
- Significantly reduce costs while maintaining quality
This is especially critical for high-frequency Agents, enabling substantial operational cost reduction without compromising user experience.
## Paper Information
- Title: RSI-Router: Evolving Subtask-Level LLM Routing and Skills for Cost-Efficient Agents
- arXiv: 2609.34712
- Publication Date: September 2026
- Research Areas: Model routing, cost optimization, Agent systems