AI Agent best practices

Use these practices to keep Agents accurate, affordable, and safe in production.

​

Prompts

• Lead with role → constraints → output format
• Ban inventing prices, legal advice, or competitor claims
• Force language mirroring (“reply in the user’s language”)
• Include a concrete JSON example when you need rich responses
• Inject contact fields carefully; handle empties

​

Knowledge and files

• Prefer focused sources over one mega-document
• Remove low-value crawled pages after indexing
• Re-train after major content changes
• Put immutable rules in the prompt, bulky facts in knowledge/files

​

Tools (functions & MCP)

• Name tools clearly (get_order_status)
• Describe when to call them, not only what they return
• Validate all tool arguments server-side
• Prefer idempotent writes
• Scope MCP credentials to least privilege

​

Routing and fallbacks

• Keywords for hard overrides; Agent for natural language — priority
• Always define human handover
• Pair Agents with Flows when steps must be exact

​

Cost and quality

​

Release checklist

1. Provider connected and model selected
2. Prompt constraints reviewed by a human
3. Knowledge/files answer the top 20 real questions
4. Tools tested for success and failure
5. Handover + fallback Flow verified
6. Channel-specific rich messages tested on device

​

Next