Prompt Engineering for AI Agents: Patterns That Work

Prompting an agent is different from prompting a chatbot. The model runs a loop, calls tools, and makes decisions — so your prompt needs rules for planning, tool use, and stopping. This is prompt engineering for agents.

The agent system prompt

1. Role: what the agent is and what it must never do.
2. Process: plan first, then act; use tools to verify.
3. Tool rules: when to call each tool, what to avoid.
4. Stop conditions: when to finish or ask a human.
5. Format: how to report results and failures.

Tool-call patterns

• Decide before you call: state what you're checking and why.
• One action per call — avoid compound tools when possible.
• Read results before acting on them.
• Retry differently after a failure, not identically.
• Escalate to a human after N failed attempts.

Iteration and stop rules

Agents can loop. Encode stop conditions in the prompt and enforce them in code: max steps, token budget, no-op detection, and timeout. The prompt sets intent; the code sets hard limits.

Common agent prompt failures

• Vague goals — the agent wanders.
• No tool priority — it calls the wrong tool.
• No stop rule — it loops or keeps going.
• Ignoring failures — it repeats the same action.
• Overconfident conclusions — it doesn't verify with tools.

Evaluate agent runs

FAQ