Connect
Guardrails
Guardrails are runtime checks on an agent’s input and output. The Guardrails page shows which guardrails are on for every agent in your workspace, and lets you turn them on or off in one place. Open it from Connect → Guardrails in the left menu.
Guardrails are off by default and set separately for each agent.
1.The guardrails
| Guardrail | What it checks | Modes |
|---|---|---|
| PII redaction | Personal data, such as email addresses, phone numbers and SSNs, in both the input the agent receives and the answer it returns | Redact masks the data, Warn flags it, Block stops the message |
| Groundedness | Whether the answer is supported by its sources | Warn flags the answer, Block withholds it |
| Reflection | Runs one self-review pass before the agent replies | On or off |
2.See your coverage
The page opens on the Coverage tab. The counts at the top show:
| Count | What it counts |
|---|---|
| Agents | All the agents in your workspace |
| PII redaction on | Agents with PII redaction turned on |
| Groundedness on | Agents with groundedness checks turned on |
| Reflection on | Agents with reflection turned on |
| No guardrails at all | Agents with no guardrail turned on. This count is highlighted, because these agents are the ones to check first. |
Below the counts, the table lists every agent with a switch for each guardrail. Type in Search agents to find an agent; the number beside the search box, such as 14 of 14, shows how many agents match.
3.Turn guardrails on
Find the agent in the table.
Turn on the switch under PII redaction, Groundedness or Reflection.
For PII redaction and Groundedness, choose the mode that appears under the switch: Redact, Warn or Block for PII redaction, and Warn or Block for Groundedness.
The change applies to the agent straight away. To turn a guardrail off, turn its switch off.
What happens when a guardrail is triggered
For PII redaction, the mode decides what the agent does when it finds personal data:
- Redact: The personal data is masked, and the agent continues.
- Warn: The agent continues, and a message in the chat asks for the personal data to be redacted, such as “Please first redact all sensitive personal…”.
- Block: The run stops, and the chat shows an error saying what kinds of personal data were found, such as “Blocked by this agent’s PII guardrail (set to Block): the message appears to contain personal information (2 email, 1 credit card, 3 phone).”
For Groundedness:
- Warn: The run completes and the user still gets the answer. A warning is recorded with the run, such as “Groundedness check skipped — the agent retrieved no knowledge passages.” Use Warn to keep an eye on unsupported answers without interrupting users.
- Block: If the answer can’t be verified against the agent’s sources, it doesn’t reach the user or the next blocks. The run stops at the agent, and the chat shows a message saying the response couldn’t be verified.
4.Use a preset
Select Presets beside an agent’s name and pick one to set all of that agent’s guardrails at once:
| Preset | What it turns on |
|---|---|
| Healthcare | Blocks on PII, blocks ungrounded claims, and reflects before replying. Patient data never gets through. |
| Finance | Masks account and card numbers and continues, flags unsupported claims, and reflects before replying |
| Balanced | Warns on PII and ungrounded claims without blocking either, with no reflection |
| Clear all | Turns every guardrail on the agent off |
You can still change any switch or mode after applying a preset.
5.See guardrail activity
The Activity tab shows every guardrail check that has run, whether it caught something or not. Choose the period at the top right: 7d, 30d or 90d.
A hit is a check that caught something. The counts at the top show:
| Count | What it shows |
|---|---|
| Checks run | How many guardrail checks ran in the period |
| Hit rate | The share of checks that were hits |
| PII redaction, Grounded, Reflection, Max length, JSON shape | For each kind of check, the number of hits and the number of checks, such as 5/40 |
Below the counts:
- By agent: For each agent, the number of hits, the number of checks, and the hit rate.
- Recent checks: The latest checks, newest first, with the agent, the status of each check and the time. Expand a check to see its details.
An agent with a high hit rate is worth a closer look: open its runs in Traces to see what the guardrails caught.