Agent & Orchestrator Prompts

Agents are models that can take actions: call APIs, run searches, or return results. When you design agent prompts, be explicit.

What to include in an agent prompt:

• Tools list: Describe each tool the agent can call, its inputs, outputs, and example calls.
• Calling format: Show exactly how to format tool calls and how the agent should present results.
• Limits and safety: Give clear stop conditions (e.g., max calls, timeouts) and how to handle failures.

Example tool description (simple):

“Tool: search(query) → returns a list of titles and short snippets. Call like: CALL search('latest AI models') and then pick the most relevant snippet for the user.”

Memory and context (short):

• Keep memory small and useful: store short summaries, user preferences, and recent important events.
• Summarize long histories before sending to the model; never dump raw logs with secrets.
• Decide what to persist (important facts) and how long to keep it.

Safety tips for agents:

• Validate outputs from tools (check types, required fields) before using them in later steps.
• For high-risk actions, require human approval or an explicit confirmation step.
• Log actions and decisions (with redaction for privacy) so you can audit agent behavior.

Designing agents well makes them reliable helpers rather than unpredictable actors.