A maker builds an agent that calls tools to look up records for the end user. The session trace shows the tools running fine and the model producing a real answer — but the person chatting with the agent only ever sees a generic "can't answer" (or blank) response.
Symptom
In the session trace, everything upstream looks healthy: the agent calls its tools, gets valid results back, and the model produces what looks like a complete final answer. But the response the end user actually receives is either empty or a generic blocked-response message — the platform's default reads "This query cannot be answered due to content guidelines. Can you please rephrase your request and try again?", though your agent may have a custom version of that message.
Checking the agent's Guardrails — PII Masking, Content Filters, Custom Word Filter, Denied Topics — everything there appears turned off, which makes the block look impossible to explain.
Cause
Guardrails has a section most people don't think to check: Hallucinations. It runs two independent checks against the model's response — Grounding (how well the answer is supported by the reference content it was given) and Relevance (how relevant the answer is to the user's question) — each scored 0–10 by a judge model, with a threshold you set.
These are separate from Content Filters, PII Masking, Denied Topics, and the Custom Word Filter, so disabling those other guardrails does nothing to Hallucinations — it has to be turned off (or its threshold adjusted) on its own. Grounding is judged against the whole conversation trace, including tool calls and tool results, not just retrieved knowledge, so a tool-based answer is still graded rather than skipped. But if the judge model's scoring output can't be parsed for any reason, the score defaults to 0 rather than passing the check, which falls below almost any threshold and blocks the response — even though the underlying answer was correct.
Fix
Open the agent. Under Guardrails in the agent's side menu, open Hallucinations. It's the second item, just below PII Masking.
If your agent doesn't need groundedness checking, turn Grounding off. If you do need it, lower the Grounding Score Threshold so valid answers stop scoring below it, and confirm the Hallucination Detection Model is configured correctly — a judge whose scoring output can't be parsed defaults to a 0 score and blocks the response.
Do the same review for Relevance if it's enabled and could be affected the same way.
If the agent's Hallucinations checks are already off, also check AI Setup > Guardrails > Hallucinations. Workspace-level settings apply to any agent that hasn't set its own value.
Changes in Guardrails save automatically — no separate Save step is needed.
To verify, re-run the query that previously returned the generic blocked message. The agent should now return its real answer. In the session trace, confirm the response no longer shows a Hallucinations block, and that a groundedness/relevance score (if still shown) no longer falls below your threshold.
Prevention
"All guardrails are off" usually means all the guardrails you remembered to check. Hallucinations (Grounding and Relevance) is a separate check that blocks responses independently of Content Filters, PII Masking, Denied Topics, and the Custom Word Filter — so when tools succeed and the model produces a real answer but the user still gets a blocked/generic response, check Hallucinations specifically, on the agent and in AI Setup, rather than re-checking the other four.
Note: A judge model whose scoring output can't be parsed silently defaults to a score of 0 — this looks like a real hallucination failure but is actually a scoring/parsing problem with the Hallucination Detection Model itself.