# Approval Architecture **Domain:** Artificial Intelligence / Delegated Authority **Doc Type:** Concept Node **Maturity:** Developed ## Definition **Approval architecture** is the system of permissions, checkpoints, receipts, revocation controls, and takeover paths that determines when an AI agent may act independently and when authority must return to a person. For agentic systems, approval is not an ornamental confirmation dialog. It is the boundary between reversible delegated work and consequential state change. The design must answer: who is the principal, what has been delegated, which data and tools are available, what action is about to occur, what evidence supports it, and how the authority can be withdrawn. ## Reversibility Gradient - **usually delegated:** research, comparison, drafting, organizing, monitoring, and simulation; - **conditionally delegated:** form filling, scheduling, account changes, and communications that remain reviewable before release; - **normally approval-gated:** sending messages, purchases, bookings, payments, credential changes, publication, legal or financial commitments, and disclosure of sensitive information; - **human takeover:** authentication, identity proofing, ambiguous consent, conflict among constraints, or an exception the agent cannot resolve safely. The exact boundary depends on the person and task. What must remain constant is legibility: the person should know whether a model, another agent, a connected service, or a human contractor will execute the action. ## Product Examples [[wiki/Meta Muse|Muse]] describes approvals before actions such as sending emails and making purchases, together with an audit trail and scoped connections. [[wiki/Grok Bot|Grok Bot]] treats the agent computer as a workspace that can be previewed or taken over. [[wiki/OpenClaw|OpenClaw]] exposes operator-controlled gateway, pairing, sandboxing, tools, and policies. These are different implementations of the same constitutional question: where does delegated machine action end and the principal's authority resume? ## Relationships [[wiki/Agentic AI|Agentic AI]] · [[wiki/Personal AI Agent|Personal AI Agent]] · [[wiki/Human-in-the-Loop|Human-in-the-Loop]] · [[wiki/Delegated Verification|Delegated Verification]] · [[wiki/Provenance|Provenance]] · [[wiki/Agent Identity|Agent Identity]] ## Sources / Provenance - [Anthropic — Building effective agents](https://www.anthropic.com/engineering/building-effective-agents) - [OpenClaw — project architecture and security boundary](https://github.com/openclaw/openclaw) - [xAI — Designing Grok Bot for a world of persistent agents](https://x.ai/news/designing-grok-bot) - [Meta — Introducing Muse](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/) - [[wired/Wired About Personal Agents Named M|Wired About Personal Agents Named M]] **As of:** 2026-09-28