Skip to main content

Reasoning Configuration Guide

This guide explains how to configure reasoning effort for AI models in Radium.

Configuration Methods​

1. CLI Flag (Highest Priority)​

Override reasoning effort for a single execution:

rad step my-agent "Prompt" --reasoning high

Options:

  • --reasoning low: Minimal reasoning
  • --reasoning medium: Standard reasoning (default)
  • --reasoning high: Maximum reasoning

2. Agent Configuration​

Set default reasoning effort in agent TOML file:

[agent]
id = "my-agent"
name = "My Agent"
reasoning_effort = "high"  # low, medium, or high

This becomes the default for all executions unless overridden by CLI flag.

3. Persona Configuration​

Configure reasoning through performance profiles:

[agent.persona]
performance.profile = "thinking"  # speed, balanced, thinking, or expert

Performance profiles map to reasoning capabilities:

  • speed: Fast models, lower reasoning
  • balanced: Balanced speed and quality (default)
  • thinking: Optimized for deep reasoning
  • expert: Expert-level reasoning, highest cost

Precedence Chain​

Reasoning effort is resolved in this order:

  1. CLI flag (--reasoning) - Highest priority
  2. Agent config (reasoning_effort in TOML)
  3. Default (medium) - Lowest priority

Examples​

Example 1: Agent with High Reasoning​

[agent]
id = "math-solver"
name = "Math Problem Solver"
reasoning_effort = "high"
engine = "gemini"
model = "gemini-2.0-flash-thinking"

Usage:

rad step math-solver "Solve: x^2 + 5x + 6 = 0"
# Uses high reasoning from config

Example 2: CLI Override​

[agent]
id = "general-agent"
reasoning_effort = "low"  # Default to low

Usage:

rad step general-agent "Complex problem" --reasoning high
# CLI flag overrides config, uses high reasoning

Example 3: Default Behavior​

[agent]
id = "simple-agent"
# No reasoning_effort specified

Usage:

rad step simple-agent "Simple question"
# Uses default (medium) reasoning

Provider-Specific Configuration​

Gemini​

Gemini thinking models (e.g., gemini-2.0-flash-thinking) support thinking mode:

[agent]
id = "gemini-thinking-agent"
engine = "gemini"
model = "gemini-2.0-flash-thinking"
reasoning_effort = "high"

The reasoning effort maps to Gemini's thinkingConfig.thinking_budget:

  • Low: 0.3 (minimal thinking)
  • Medium: 0.6 (standard thinking)
  • High: 1.0 (maximum thinking)

Claude​

Claude models support extended thinking:

[agent]
id = "claude-thinking-agent"
engine = "claude"
model = "claude-3-opus"
reasoning_effort = "high"

The reasoning effort maps to Claude's thinking.thinking_budget:

  • Low: 0.3 (minimal extended thinking)
  • Medium: 0.6 (standard extended thinking)
  • High: 1.0 (maximum extended thinking)

Cost Considerations​

Reasoning effort directly impacts cost:

ModelStandardThinking (High)Multiplier
Gemini Flash Exp$0.075/$0.30--
Gemini Flash Thinking-$0.20/$0.80~2.7x
Claude Sonnet$3.00/$15.00$3.00/$15.00*~1.0x*
Claude Opus$15.00/$75.00$15.00/$75.00*~1.0x*

*Claude models use extended thinking which increases token usage, effectively increasing cost per request.

Best Practices​

  1. Use appropriate reasoning levels:

    • Simple tasks: low or medium
    • Complex problems: high

  2. Monitor costs:

    • Check token usage with --show-metadata
    • Use thinking models only when needed

  3. Combine with model selection:

    • Use thinking models for complex tasks
    • Use standard models for simple tasks

  4. Test reasoning levels:

    • Start with medium and adjust based on results
    • Use high only when necessary

Troubleshooting​

Reasoning effort not taking effect​

  • Verify model supports thinking mode (check model name)
  • Check precedence: CLI flag overrides config
  • Ensure reasoning effort is spelled correctly (low, medium, high)

Unexpected costs​

  • Reduce reasoning effort level
  • Use standard models for simple tasks
  • Monitor token usage in metadata

Slow performance​

  • Lower reasoning effort for faster responses
  • Use streaming mode for real-time output
  • Consider using faster models for time-sensitive tasks

See Also​