Governance
Responsible AI in Studio
Studio’s Governance pages keep agents safe and accountable from the first draft to production: you classify each agent’s risk, get an admin’s sign-off when it is high, test it against adversarial users, approve risky actions as they happen, and keep a record of every change and run.
1.The Governance pages
You’ll find these pages under Governance in the left menu.
| Page | What it’s for |
|---|---|
| Responsible AI | Classify each agent’s risk, assign an oversight owner, and record an admin’s sign-off for high-risk agents |
| Audit Log | See every change to an agent’s configuration: what changed, who changed it, and whether it changed what the agent can do |
| Agent Simulation Engine | Test an agent against invented adversarial users and get a judged pass-rate report |
| Approval Flows | Approve or reject the requests agents and workflows send for human approval |
| Traces | Follow every run step by step, and track success rate, cost and latency |
| Dashboards | Build KPI boards from your data. Each KPI card is built and kept up to date by its own agent, and you can share a board with a view-only link |
2.How governance fits an agent’s life
Each Governance page covers one stage, from building an agent to watching it in production.
Build the agent with guardrails turned on, and add safety blocks such as PII Redact and Human Approval to its workflow.
Test it in the Agent Simulation Engine, and fix the gaps the report finds.
Classify its risk in Responsible AI and assign an oversight owner. A High-risk agent can’t be deployed until an admin signs it off; a Prohibited agent can’t be deployed at all.
Deploy the agent.
Approve risky actions as they happen in Approval Flows.
3.Safety checks on agents
Guardrails are runtime checks on an agent’s input and output. You can turn them on for any agent, whether you built it on the canvas or with New agent. They are off by default and set separately for each agent.
| 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 |
The Guardrails page shows which guardrails are on for every agent, and how many agents have none at all, so you can switch them on across your agents in one place.
Agents can also use Clarify first, which asks the user questions before answering a request that is unclear. Studio also filters harmful content and attempts to override an agent’s instructions.
When an agent answers from a knowledge base, users can see the citations for its answer.
4.Workflow blocks that support Responsible AI
Four blocks add safety checks inside a workflow. Each is described in full in the Blocks reference.
| Block | Use it to | Place it |
|---|---|---|
| PII Redact | Mask emails, phone numbers, SSNs, credit card numbers and IP addresses | Before an AI block, an output, or a message that should not see personal data |
| Validator | Check inputs against rules, such as required fields, formats and allowed values | Where users type free text or form data |
| Evaluator | Grade an answer against your criteria and a pass threshold from 0 to 100 | After an AI Agent, before its answer is sent |
| Human Approval | Pause the workflow until a person approves the next step in Approval Flows | Before an action that is hard to undo, such as sending an email or changing data |