AI · Safety
AI Guardrails on VegaDūta
Guardrails on VegaDūta are a platform layer, not something you prompt into each agent and hope holds. Built-in detectors cover PII, toxicity, profanity, jailbreak attempts, and policy violations; a configurable rule engine and an optional external safety classifier extend them; and human approval gates sit in front of sensitive tool calls so an agent asks before it acts.
Everything is audit-logged, and flagged content lands in a moderation review queue where a person makes the final call.
Detection built in
Agent traffic passes through detectors for personally identifiable information, toxicity, profanity, jailbreak attempts, and policy violations. You do not assemble these from scratch per agent — they are platform capabilities you configure.
- PII detection to keep personal data from leaking through agent responses
- Toxicity and profanity screening on conversations
- Jailbreak detection for prompt-injection and manipulation attempts
- Policy violation rules for tenant-specific red lines
Configurable rules and an external classifier
The rule engine lets you define what counts as a violation for your business, beyond the stock detectors. An external safety classifier can be plugged in as an additional judgment layer when the built-in checks are not enough on their own.
Humans stay in the loop
Sensitive tool calls can require explicit human approval: the agent proposes, a person disposes. Approvers can apply policy overrides where appropriate and retry actions that failed, and flagged content flows into a moderation review queue rather than being silently dropped. Every decision is captured in the audit log.
Frequently asked questions
What guardrails does VegaDūta include out of the box?
PII detection, toxicity screening, profanity filtering, jailbreak detection, and policy violation rules, plus a configurable rule engine to add your own. An external safety classifier can be layered on top, and a moderation review queue holds flagged content for human review.
Can I require approval before an agent takes an action?
Yes. Sensitive tool calls can be gated behind human approval gates — the agent requests the action and waits for a person to approve or reject it. Approvers can also apply policy overrides and retry failed actions.
How do I know what my agents have been doing?
Guardrail decisions and approvals are audit-logged, and the platform's Langfuse-based tracing records agent runs end to end. Between the two you can reconstruct what an agent did, what was blocked, and who approved what.
Can I customize the rules for my business?
Yes. The rule engine is configurable per tenant, so you define policy violations that match your domain on top of the stock detectors. Approval gates are likewise configured per tool, so you choose which actions need a human sign-off.
See it working in two minutes
The sandbox provisions a real tenant — describe an agent in one sentence and test it, no account, no card. Or browse ~90 industry workflow recipes to see what teams build.
