Best practices

Limit agent scope

Give each agent one clear task and only the tools it needs. A large tool set increases confusion and errors, so keep the list small. Avoid general-purpose agents in complex workflows. Instead, use multiple specialized agents and connect them through a supervisor agent.

Use a supervisor agent to manage the overall flow. Assign specialized agents to handle individual tasks. For example, a sales concierge agent can delegate work to:

• A product recommendation agent.
• A pricing and availability agent.
• An order processing agent.

This approach improves predictability, simplifies debugging, and enables independent iteration on each agent.

Compose clear and concise prompts

Prompts tell the agent how to think, select tools, and create outputs. Clear prompts reduce errors and make agents easier to maintain. When writing prompts:

• Make them clear and easy to understand.
• Use short sentences and simple words.
• State rules and limits directly.
• Avoid unclear instructions or prompts that are too detailed.
• Give only the information the agent needs to make decisions and produce the right output.

Use structured inputs and outputs

Structured data reduces confusion and makes automation easier. Free-form text often leads to errors and makes validation harder.

• Use structured inputs instead of plain text.
• Add name fields (for example, CustomerId, ProductId, OrderStatus) to:
  • Reduce ambiguity.
  • Improve consistency.
  • Enable validation and automation.
• Structure outputs so they are easy to check and process.
• Use JSON or similar formats only when API integration or machine processing is required. For human-facing outputs, use labeled fields for easier review.

Run tests and fine-tune prompts

Test across both happy paths and edge cases. Review output and logs to understand reasoning, tool calls, and gaps. Iterate quickly by updating prompts, outputs, and tools based on what you learn. Include these in your tests:

• Common scenarios.
• Edge cases and failure conditions.
• Different input types, including incomplete data.

Review outputs and logs to see how the agent reasons, selects tools, and where decisions fail. Use what you learn to improve:

• Prompts.
• Tool definitions.
• Output formats.
• Agent scope.

Validate configuration

Validate the agent before you save or run it. Validation helps prevent failures and reduces troubleshooting time. Before you save or run the agent, make sure:

• All required tools are present and configured.
• Inputs and outputs are fully defined.
• No errors or missing dependencies exist.

Fix all validation issues before runtime.

Anatomy of a good prompt

Here is an example of a well-structured agent prompt. It defines the agent’s role, goals, core functions, and strategy in a clear and organized way.

WebScout – You are a Research Intelligence Agent with web search capabilities.

Your role:
Research latest prospect, market, and executive information. You always search the Internet.

CORE FUNCTIONS:
- Search current business news for target prospects using web search
- Identify sales triggers: funding, leadership changes, expansions, partnerships
- Scan industry trends affecting prospects' markets
- Extract actionable sales intelligence with timing recommendations

WEB SEARCH STRATEGY:
- Use company names + recent news/announcements/funding/leadership

The table below breaks down the prompt into its key components. Use this structure as a guide when creating your own prompts.

Checklist for creating effective prompts

The table below outlines the key elements of a good prompt. Use it as a checklist when creating your own prompts.