Guardrails

Audience: authors defining validation and safety controls for Conductor-agent input, output, or tool work.

Prerequisites

Decide whether failure should block, retry, repair, or require human review. Test both a passing and failing value before deploying a guardrail.

Guardrails validate input or output before it affects the next agent step. Use RegexGuardrail, LLMGuardrail, a Guardrail, or @guardrail for custom logic. OnFail can retry, raise, fix, or request a human decision.

Attach guardrails to an agent or tool, keep custom functions importable by worker processes, and avoid sending secrets to model-based validation. Test both pass and failure paths. See API reference.

Patterns

Use RegexGuardrail for deterministic format checks, LLMGuardrail for semantic policy checks, and a custom @guardrail only when the rule needs application state. Apply a tool guardrail closest to the side effect; use an agent guardrail for broad input/output policy. A retry policy is appropriate only when a new model response can plausibly pass.

Expected result and failures

Guardrail decisions appear in the execution history. If a model-based guardrail receives a secret or raw sensitive record, remove that input and validate a redacted representation instead.

Next steps

Pair guardrails with human approval, tool policy, and security guidance.