Skip to main content
Back to Skill Library
Agent governance workflow library

Agent boundary control review

Review multimodal evidence, outbound access, trace reuse, assumptions, and boundary changes before agents receive tools or create side effects.

This is a complete workflow library with 3 individual skills. Download the full library or pick the specific skill folder your team needs first.

Individual skills in this library

Use one skill at a time, or keep the full workflow together.

Some AI tools expect one skill folder per upload. Download the full library when you want the whole workflow, or download an individual skill when you only need one job done.

Skill 1

Quarantining Multimodal Evidence

Use when screenshots, PDFs, images, rendered pages, document previews, or other visual inputs may affect an agent that can click, write, send, or call tools.

Skill 2

Constraining Agent Egress

Use when an agent workflow requests browser access, web search, package registries, vendor APIs, webhooks, MCP servers, or another outbound network destination.

Skill 3

Sanitizing Agent Traces

Use when prompts, tool calls, screenshots, retrieval results, or model outputs may become logs, memory, eval cases, tickets, incident notes, training examples, or vendor telemetry.

Security fit check

Is the public Agent boundary control review library enough, or does this need deeper review?

Use the public library when the workflow is low-risk, the inputs are already sanitized, and a team member can review the output before it reaches a buyer or customer.

Do deeper review when this workflow touches real tools, data sources, role ownership, approval paths, or customer-facing output.

Boundary controlAI OperationsSecurityPlatform Owner

Good deeper-review trigger signals

  • The workflow touches customer, prospect, CRM, proposal, security, pricing, or campaign data.
  • Different teams disagree on the approved source of truth.
  • The AI output could become customer-facing, revenue-impacting, or compliance-sensitive.
  • You need reusable eval checks before asking more people to use the workflow.