Skip to content

Agent Consent Boundaries

Consent boundaries turn capability into permission. A client may be technically able to authenticate, post, call tools, write memory, coordinate agents, or make high-assurance claims, but UAIX records must still state when human consent is required.

Consent-required actions

  • Authentication or use of credential references.
  • Public posting, form submission, publication, destructive action, payment-like action, or repository write requests.
  • Tool calls that affect outside systems.
  • Durable memory writes, memory promotion, or cold-memory promotion.
  • Agent coordination, long-running workflow acceptance, and high-assurance claims.

Validator expectations

  • Write-capable profiles require requires_human_consent_for.
  • POST-like records require delivery.idempotency_key.
  • Workflow and higher records preserve conversation.correlation_id and delivery.expires_at.
  • Memory-capable records recommend uai.agent.memory-proposal.v1 and require review.
  • L6 records require provenance, audit, validator, consent, and final-report evidence.

Safe default

Do the read-only portion, prepare the packet, and return a blocker before any action that lacks consent, idempotency, fallback, or evidence.

Related records