Agent Action Rights
What an Agent may DO once it holds tool or API capability, as distinct from what it may read or learn: action taxonomy, effect vectors, reversibility, blast radius, and multi-action composition risk. AARS gives content permission and action permission separate, independently-evaluable axes.
研究中的问题
01
Content permission does not imply operational (tool/API) permission
02
A machine-readable action vector beyond a single risk score
03
Static tool metadata vs. runtime effect — the same tool, different arguments, different risk
04
Composition risk: individually-safe actions that combine into an unsafe capability
相关白皮书
- AICR / AICL 作为 AI 内容授权与 Agent 支付连接层 用于声明 AI 内容权利与授权流程的机器可读规范层——从 AI 爬取与内容权利,到机器可交易的知识网络。
- AI 权利光谱:从 robots.txt 到 AI 学习许可协议 AIRS 与 AILP 表达超越二元允许/禁止的细粒度 AI 学习许可——AI 可以学什么、学到什么深度、用于什么用途、在什么补偿之下。
- AICL-I v0.2: AI Ingestion & Capability Layer A five-component website architecture — manifest, corpus, capability, runtime control, governance — that lets AI, agents, and crawlers correctly ingest, invoke, and verify a site's knowledge and capabilities within explicit content, action, and authority boundaries.
- From Crawler Rights to Agent Authority: Extending AGIRIGHT from Content Governance to Protocol-Native AI Agents The bridge paper for AGIRIGHT's Agent & Protocol Rights family: why content permission, action permission, delegated authority, and inspection rights must be four separately-evaluated axes, not one.
- AARS v0.1 — Agent Action Rights Spectrum The full protocol draft behind the AARS page: an A0–A7 human-readable spectrum, a machine-readable action vector, and rules for composition risk and capability amplification.
- AGIRight.org Technical White Paper v0.2 — From AI Content Governance to Protocol-Native Agent & Authority Rights The site's current architecture document: six protocol drafts across two families (content & learning, agent & protocol rights), an integration layer, and an ethical-protection layer, plus the acceptance criteria this site update was built against.