Validate inputs and final outputs in an agent workflow

Goal

Prevent inputs or answers that violate application rules from proceeding.

Steps

1. Separate structural schema validation from business/policy checks: the former constrains field shape, the latter judges whether a request or result is allowed. [openai-guardrails]
2. Use @input_guardrail for the first input and @output_guardrail for final output; return GuardrailFunctionOutput(..., tripwire_triggered=True) to block. [openai-guardrails]
3. For policy checks on every custom function-tool call, attach @tool_input_guardrail / @tool_output_guardrail rather than relying on agent-level checks. [openai-guardrails]
4. Add tool-level checks to each side-effecting custom tool; do not assume agent-level checks cover handoffs. [openai-guardrails]
5. Trigger each check with allowed and rejected cases and confirm failures stop at the intended boundary. [openai-guardrails]

How to verify

Record trigger tests and blocked execution points for each guardrail; verify handoff and tool boundaries are covered.

Troubleshooting

If a check does not run, verify which agent or tool owns that workflow boundary; handoffs and hosted tools may follow different paths.