Your agent needs a temperature schedule (or it'll be creative when you need precision)

Your agent is using the same temperature for everything. That's why it writes poetry when you ask for JSON, and outputs robotic responses when you need creative marketing copy.

Temperature controls randomness. 0.0 is deterministic, 1.0 is chaotic. Most people set it once and forget it. But your agent does different types of work that need different levels of creativity.

Here's the temperature schedule I use for my coding agent:

TASK_TEMPERATURES = {
    "code_review": 0.1,     # Be precise, catch real issues
    "bug_fixes": 0.2,       # Systematic debugging
    "documentation": 0.3,   # Clear but not robotic
    "refactoring": 0.4,     # Some creative restructuring
    "feature_planning": 0.6, # Creative problem solving
    "naming_things": 0.8,   # Actually be creative here
}

The pattern is simple: precision tasks get low temperatures, creative tasks get high temperatures.

For my business agent, the schedule looks different:

BUSINESS_TEMPERATURES = {
    "financial_analysis": 0.1,  # No creativity with numbers
    "meeting_summaries": 0.2,   # Accurate, not embellished
    "email_responses": 0.4,     # Professional but human
    "content_ideas": 0.7,       # Creative brainstorming
    "marketing_copy": 0.8,      # Maximum creativity
}

The implementation depends on your setup. If you're using the OpenAI API directly, you can route by task type:

def get_temperature_for_task(task_type):
    return TASK_TEMPERATURES.get(task_type, 0.5)  # Default fallback

# In your agent loop
temperature = get_temperature_for_task(current_task)
response = client.chat.completions.create(
    model="gpt-4",
    temperature=temperature,
    messages=messages
)

If you're using Claude Desktop or another interface, you can't control temperature directly. But you can prompt for it:

For code reviews: "Be precise and systematic. Focus on actual issues, not style preferences."

For creative tasks: "Be creative and explore multiple approaches. Don't just give me the obvious solution."

Pro tip: Log your temperature settings with task outcomes. You'll quickly see which tasks need adjustment. My bug-fixing temperature started at 0.1 but I bumped it to 0.2 because the agent was missing creative debugging approaches.

The biggest mistake I see is using high temperature for everything because "creativity is good." But when your agent is generating API keys or processing financial data, creativity kills reliability.

Start with these ranges:

• 0.0-0.2: Data processing, code review, calculations
• 0.3-0.5: Documentation, email responses, summaries
• 0.6-0.8: Content creation, brainstorming, marketing
• 0.9-1.0: Rarely useful (too chaotic)

Temperature scheduling isn't just about output quality — it's about trust. When your agent is precise for precise tasks and creative for creative tasks, you stop second-guessing every response.

If you're running coding sessions that need this kind of task-aware temperature control, you need a system that can handle different types of work with different configurations.