What: A new system that allows Claude to dynamically adjust how much reasoning it applies to tasks based on an experiment-controlled effort level.
How it works: This feature is controlled by the Statsig experiment flag tengu_effort_exp with parameter tengu_effort_level. When enabled, it accepts three levels:
Details:
- The feature automatically injects reasoning effort instructions into Claude's system prompt when the experiment is active
- Lower effort values prioritize faster, more efficient responses
- Higher effort values enable maximum reasoning and thoroughness
- The effort scale ranges from 0-100, with Claude instructed to vary reasoning depth accordingly
- This is an A/B test feature - only users enrolled in the Statsig experiment will have access
- Evidence:
function sQI() at line 470523,aQI mapping at line 470940,system prompt integration at line 470717
Technical implementation:
- New function
sQI()reads experiment configuration and generates reasoning effort instructions - Mapping object
aQI = { low: 45, medium: 75, high: 99 }converts string levels to numeric values - The generated
<reasoning_effort>XML tag is appended to the system prompt array - Experiment values are retrieved via the Statsig client using the
d3()helper function