Write Effective AI Agent Prompts

TL;DR

Write clear, specific AI Agent prompts using natural language and concrete examples, define fallback behaviors for edge cases, and always test in the Playground before deploying. Start small, avoid complex multi-step logic, and chain agents (one triggers another via field changes) for dependent workflows. Use multi-action agents for simultaneous updates on the same task, and leverage calculated fields to give agents data they can't access directly (like capacity or metrics).

• Overview
• Do's for AI Agent Prompts
• Don'ts for AI Agent Prompts
• Optimization Tips
• AI Agent Chaining
• Chaining vs. Multi-Action
• Use Calculated Fields to Extend What AI Agents Can See
• What’s Next?

Overview

Well-written prompts are the key to reliable agent behaviour. Use these do's, don'ts, and patterns to get consistent results.

Do's for AI Agent Prompts

• Be specific about what you want the AI agent to look for.
• Give clear examples of the classifications or actions you expect.
• Define fallback behaviours for any edge cases.
• Use natural language - write as if you're explaining it to a helpful colleague.
• Test your prompt thoroughly in the Testing Playground before deploying.

Don'ts for AI Agent Prompts

• Don't create overly complex, multi-step logic.
• Don't assume the AI agent understands your company's terminology without explanation.
• Don't forget to address situations with unclear or incomplete information.
• Don't deploy your AI agent without testing it first.

Optimization Tips

• Start small: Test one AI agent type in a single folder before expanding.
• Test first: Always validate behaviour in the Playground before deploying.
• Combine agents: Use the Intake Agent and Triaging Agent together for complete request handling.
• Review often: Regularly refine prompts based on activity logs and team feedback.
• Communicate: Keep your team informed when AI agents are active and what's being automated.

AI Agent Chaining

One agent's action can trigger another agent. This is the recommended approach for workflows where one step depends on the result of a previous step.

How it works: Agent 1 performs an action (e.g., sets a custom field value) → Agent 2 is configured to trigger on that field change → Agent 2 performs the next step.

Examples:

• Triage → Assignment: A triage agent sets the Priority field → an assignment agent triggers on Priority changes and assigns based on the new value.
• Intake → Routing: An intake agent validates a request and sets status to "Complete" → a routing agent triggers on that status change and moves the item to the correct folder.

• Categorization → Notification: A categorization agent sets the Type field → a notification agent triggers on Type changes and posts a targeted comment.

Chaining vs. Multi-Action

Use Calculated Fields to Extend What AI Agents Can See

AI agents cannot access workload metrics, capacity data, or perform reliable arithmetic. But agents can read calculated/formula fields. Create a formula that computes the value you need (e.g., available capacity, overdue percentage, risk score), and the agent can read the result and act on it.

What’s Next?

• AI Agent Limitations and Troubleshooting