Яндекс.Метрика

Guardrails

Guardrails explained

Rules and technical limits that keep an AI agent within the scope of its task: access controls, limits, filters and human approval

Guardrails are the measures that make an AI agent's behavior predictable and safe. Instructions in the prompt tell the agent what it can and can't do, but a model can break them. That's why key guardrails are enforced at the system level: the agent simply can't do anything it doesn't have permission for.

Types of guardrails:

  • Access: only the data and actions it needs, with minimal token permissions
  • Limits: number of steps, run time, request rate
  • Filters: checks on incoming requests and the agent's replies
  • Approval: a person signs off on risky actions
  • Logging: a record of the agent's actions for review

What Pachca offers:

  • A team agent sees public channels and only the private chats it's been added to
  • You can set scopes on an employee's token, like read-only access to messages
  • Administrators decide who can use the API and create integrations
  • Actions taken through the API go into the audit log on the Corporation plan
  • You can revoke a token or disable a bot, and access is cut off immediately

Learn more in the Help center

Read the article →

Related terms