Customizing agents

Most “bad agent output” is actually bad input structure. This guide shows how to get reliable results without turning the workflow into a science project.

Browse agents

1) Make the input shape stable

Prefer structured inputs over long free-form prompts.

{
  "task": "Extract fields",
  "source": "customer_email",
  "text": "...",
  "fields": ["company", "budget", "timeline"],
  "output_format": "json"
}

2) Add constraints (what NOT to do)

Good constraints are short and enforceable:

• Don’t invent facts.
• If information is missing, return null.
• Keep answers under N sentences.

3) Use examples sparingly (but correctly)

A single good example can outperform long instructions.

• Example input
• Example output
• 1–2 “edge case” examples

4) Evaluate before you automate

Create a small test set:

• 10 normal cases
• 5 edge cases
• 5 adversarial cases (ambiguous, messy, missing info)

Checklist

• Build a 20-case test set
• Define what “pass” means
• Retest after any major change

5) Handle safety and privacy

• Don’t send secrets in prompts
• Minimize PII where possible
• Log outcomes, not raw customer data

Security practices