Guardrails
Guardrails are runtime enforcement mechanisms that intercept agent actions before they execute. Where policies declare what agents may or must do, guardrails actively gate operations at execution time by blocking, modifying, or warning on attempted file writes, tool invocations, and content generation. Every interception is recorded in the audit log, which makes guardrails both an active safety layer and a compliance surface for later review.
Guardrail Classes
Guardrail Classes
The named categories of guardrails: grounding, file-write, tool-invocation, and content-generation, and how each maps to a distinct interception point in the runtime.
Guardrail Configuration
Guardrail Configuration
How to declare and scope guardrails at the org, project, task, or agent level, including precedence rules when multiple scopes overlap.
Grounding Guardrail
Grounding Guardrail
The generation-time control that checks whether factual claims in agent output trace to authoritative context sources. It fires inline during content production, before output exits the agent's execution boundary.
File Write Intercepts
File Write Intercepts
Scope enforcement for filesystem operations. Agents may only write within their own workspace; shared-surface changes route through PRs. Enforcement happens at both the filesystem layer and at commit time.
Tool Invocation Intercepts
Tool Invocation Intercepts
Pre-execution gates on tool and command use. Covers availability checks, scope validation, and parameter sanitization before a tool call proceeds.
Content Generation Intercepts
Content Generation Intercepts
Inline checks that evaluate generated content against policy constraints such as tone, format, and prohibited patterns before the output is committed to a deliverable.
Violation Handling
Violation Handling
What happens when a guardrail fires: block vs. warn vs. rewrite outcomes, how violations surface as errors or warnings, escalation to human-in-the-loop review, and the override path that requires owner-level authorization with a recorded rationale.