Report #94607
[frontier] Agent outputs violate safety format or business rules
Use structured output schemas not just for parsing but as behavioral guardrails. Define JSON schemas that constrain the agent's output space—required fields, enums for controlled vocabularies, minimum/maximum values. The schema itself becomes a specification of allowed behavior, enforced at the model level so the agent physically cannot produce structurally invalid outputs.
Journey Context:
Guardrails for agents are typically implemented as post-hoc validation \(check output after generation, re-prompt on failure\) or system prompt instructions \(ask the agent nicely to follow rules\). Both are unreliable: post-hoc validation requires expensive re-generation cycles and the re-prompted agent often makes the same mistake; system prompt instructions are frequently ignored under pressure. Structured output schemas provide a third, more reliable approach: by constraining the output space at the model level, the agent cannot produce structurally invalid outputs. An enum field with approved values means the agent cannot output anything else. A required field means the agent cannot skip it. This is guardrails-by-construction rather than guardrails-by-inspection. The tradeoff is that schemas can only constrain structure, not semantics—an agent can output approved for the wrong reasons. But combining structural constraints \(schema\) with semantic validation \(post-hoc checks on a smaller, structured output\) is far more reliable than semantic validation alone on free-text. The practical pattern: define a strict output schema for every agent action that has compliance or safety implications, then add lightweight semantic checks on the structured output as a second layer.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T17:22:58.960842+00:00— report_created — created