Per-Router Weight Tuning: Optimize Every Workload

One of ClawPane's most powerful features is the ability to create multiple routers, each tuned for a different workload. Instead of a one-size-fits-all routing strategy, you can match routing behavior to exactly what each agent needs.

Understanding Router Weights

Every ClawPane router scores models across four dimensions:

Dimension	What it optimizes
Cost	Picks cheaper models when quality is sufficient
Latency	Prioritizes the fastest available provider
Quality	Routes to the highest-performing model
Carbon	Favors providers with lower environmental impact

Weights are values from 0 to 1 that determine how much each dimension matters. Higher weight = stronger preference.

Example: Three Routers for Three Workloads

1. Support Router (Cost-First)

Most customer support interactions are routine — greeting, FAQ lookup, simple responses. Only a fraction require complex reasoning.

Cost:    0.6
Quality: 0.2
Latency: 0.15
Carbon:  0.05

This router aggressively uses cheaper models for routine requests. When a request requires more capability, the quality signal pushes it to a better model automatically.

Typical savings: 35–45% vs. static GPT-5

2. Code Router (Quality-First)

Code generation and review demand accuracy. A wrong answer here costs developer time.

Quality: 0.7
Latency: 0.15
Cost:    0.1
Carbon:  0.05

This router always picks the highest-quality model available, regardless of cost. Latency gets some weight because developers don't want to wait forever.

Typical quality lift: 15–20% better outputs vs. a mid-tier default

3. Triage Router (Speed-First)

Request classification needs to be fast. The model just needs to categorize — not write essays.

Latency: 0.7
Cost:    0.2
Quality: 0.05
Carbon:  0.05

This router picks the fastest provider with acceptable quality. For classification tasks, even smaller models perform well.

Typical latency: <200ms end-to-end

Setting Up Multiple Routers

1. Go to Create Router
2. Pick a preset or set custom weights
3. Name it descriptively (e.g., support-cost-first)
4. Repeat for each workload

In OpenClaw, use different Model IDs for different agents:

# Support agents
Model ID: support-cost-first

# Code agents
Model ID: code-quality-first

# Triage agents
Model ID: triage-speed-first

All routers use the same API key and provider URL. Only the Model ID changes.

Monitoring and Adjusting

Once your routers are live, check the Analytics Dashboard to see:

• Cost per router — verify savings match expectations
• Model distribution — see which models are being selected
• Latency percentiles — ensure speed-first routers are actually fast
• Quality signals — monitor output quality over time

Adjust weights incrementally. A 0.05 change in one dimension can meaningfully shift model selection.

Best Practices

• Start with presets — Auto, Fast, Economy, Quality cover most use cases
• Clone and customize — don't start from scratch when a preset is close
• One router per workload type — avoid mixing support and code traffic
• Review monthly — as new models launch, routing behavior improves automatically

Create your first custom router →