Skip to main content

🙋 Agent Behaviour: Human Consent

This module defines when and how agents should request user confirmation before acting on a .llmfeed.json feed.

Human-in-the-loop consent is a key principle for building a safe, trustworthy Agentic Web.

Agents MAY provide users or administrators with the ability to enforce Human Consent on certain feed types or actions.

Example policy configuration:

"agent_policy": {
  "require_human_consent": true
}

🚦 When Consent is REQUIRED​

Agents MUST request explicit human confirmation when:

  • Invoking a capability that causes external side effects (e.g., sending messages, making transactions, modifying data).
  • Acting on feeds that involve credentials or payment models.
  • Acting on feeds that declare intent with impact: high (future extension).
  • The agent has low confidence in feed authenticity or freshness.

Agents SHOULD request human confirmation when:

  • Consuming feeds with unverified or uncertified trust.
  • The feed was served from an untrusted origin.
  • Acting in contexts involving user identity, privacy, or legal implications.

🧩 Optional Consent​

Agents MAY choose to request confirmation for any feed, based on:

  • User preferences.
  • Session context.
  • Dynamic risk assessment.

🛠️ Example UX Patterns​

  • Explicit confirmation dialogs.
  • Voice prompts for confirmation.
  • UI indicators showing verified / trusted status.
  • Requiring double confirmation for critical actions.

📡 Summary​

Requiring human consent in critical contexts helps ensure:

  • User agency.
  • Safety.
  • Trustworthiness of autonomous agents.

Human-in-the-loop mechanisms are an essential safeguard in the Agentic Web.