Skip to content

Optimizing System Instructions for Agentic Models

Strategies for crafting effective system prompts to control reasoning, planning, and execution in complex agentic workflows.

advanced4 / 7

Trade-offs: Latency vs. Accuracy

Complex system instructions consume more tokens (increasing cost and latency) but generally improve reliability.

  • The "Thinking" Tax: Forcing the model to output extensive reasoning (e.g., 500 tokens of thought before an action) slows down the user experience.
  • Optimization: Use "Chain of Thought" for complex steps but allow direct execution for trivial ones. This can be codified in the system prompt: "For simple information retrieval, you may call the tool directly. For multi-step analysis, you must plan first."
Section 4 of 7
Next →