Guardrails
Guardrails define non-negotiable boundaries for how your Agent and Subagents behave.
They are critical for reliability, compliance, and user trust.
What Guardrails protect
Guardrails help prevent:
- unsafe or non-compliant responses,
- hallucinated policy claims,
- over-broad tool execution,
- poor escalation decisions,
- unintended data exposure.
Guardrail layers
Apply constraints at multiple levels:
| Layer | Scope | Best for |
|---|---|---|
| Agent-level | Global behavior | brand voice, safety baseline |
| Subagent-level | Workflow-specific | domain-specific constraints |
| Tool-level | Action boundaries | input/output execution limits |
Layering gives stronger control than relying on one prompt block.
Core Guardrail categories
| Category | Objective | Example |
|---|---|---|
| Behavioral | Response scope/tone | “Do not provide legal advice.” |
| Data | Retrieval and privacy limits | “Use only attached Drive context.” |
| Action | Tool invocation constraints | “Require confirmation before outbound call.” |
| Escalation | Human handoff policy | “Escalate after two failed clarifications.” |
Designing effective Guardrails
Use rules that are:
- explicit,
- testable,
- scenario-aware,
- conflict-free.
Prefer clear constraints over vague wording.
Example policy patterns
Safety-first response policy
- Refuse harmful requests
- Offer safe alternatives
- Keep refusal concise and respectful
Data minimization policy
- Avoid collecting sensitive data unless required
- Request only fields needed for task completion
- Mask or avoid repeating sensitive values
Escalation policy
- Ask clarifying question when uncertainty is moderate
- Escalate when confidence is low and impact is high
Validation checklist
Before deployment, verify:
- Guardrails exist at appropriate layer(s)
- High-risk prompts are tested
- Tool constraints are aligned with business policy
- Escalation paths are reachable and clear
- Subagent-specific constraints do not conflict with global constraints
Monitoring and iteration
After deployment, review:
- Threads for policy misses,
- Feedback for recurring trust issues,
- Worklist/escalation for over-trigger or under-trigger patterns.
Guardrails are living controls; refine them with real usage evidence.