Agent Spend Controls / 智能体消费控制

Definition

Agent spend controls / 智能体消费控制 are the monetary limits, eligible uses, time windows, credentials, confirmations, audit evidence, and reauthorization rules that bound how an agent spends for a user or organization.

Current Synthesis

Spend controls translate general Agent Permission Boundaries into payment authority. A useful agent needs enough budget autonomy to complete shopping, booking, token, API, data, or service tasks without asking at every step, but a stored card or broad wallet login makes scope and responsibility ambiguous.

The current synthesis is a graduated mandate: specify an amount ceiling, product or service category, merchant constraints, time window, task purpose, substitution rules, and conditions requiring renewed consent. Pair that mandate with a one-time or otherwise scoped credential, agent identity, execution trace, and dispute evidence.

Trust should determine breadth rather than disappear after one approval. Low-risk recurring tasks can receive standing limits, while high-value, regulated, identity-sensitive, or materially changed purchases require explicit confirmation. The user’s willingness to raise the maximum delegated amount is evidence of Agent Trust Calibration / 智能体信任校准, not proof that all purchases should become autonomous. The latest source also makes these controls liability evidence: they help distinguish an in-scope delegated act, a user bypassing warnings, inadequate platform design, and a product departing from its represented purpose.

Key Claims

• Budget autonomy should be sufficient for task completion but narrower than the user’s underlying payment account.
• Amount, category, merchant, duration, purpose, and substitution constraints should travel with the payment mandate.
• Repeated low-value purchases need standing permission with revocation and review, while high-risk purchases need renewed confirmation.
• Separate accounts, one-time credentials, and audit records improve attribution when spending is wrong or compromised.
• Spend limits work best when tied to authenticated principals, known agents, Verifiable Intent / 可验证意图, and enforceable execution controls.
• Compensation and dispute handling affect how much users will rationally delegate.
• Limits, warnings, confirmations, and execution records help allocate responsibility but do not predetermine the legal outcome.

Evidence

• Mandate evidence: 可以给你的 Agent 发一点零花钱了｜ S10E22 describes price, category, user-intent, merchant, and one-time-payment constraints for agent purchases.
• Self-spend evidence: 可以给你的 Agent 发一点零花钱了｜ S10E22 distinguishes consumer shopping from small budgets for tokens, APIs, reports, images, and other task resources.
• Strong-constraint evidence: 外滩大会线下圆桌｜敢把钱包交给AI吗？聊聊Agent交易爆发前夜的信任基建 says an initial 100-yuan limit must not be displaced by later task content.
• Delegation evidence: 外滩大会线下圆桌｜敢把钱包交给AI吗？聊聊Agent交易爆发前夜的信任基建 uses the maximum amount a user would delegate as a concrete test of authorization, safety, and recourse confidence.
• Liability-boundary evidence: AI Cyber Panic, Ken Griffin’s Bubble Warning, The Fall of France? distinguishes actions within delegated authority, user circumvention of warnings or terms, and an agent departing from its represented purpose.

Counterevidence & Qualifications

• Monetary ceilings do not prevent wrong-product selection, biased recommendations, privacy leakage, or prohibited non-financial actions.
• Confirmation on every purchase creates approval fatigue; standing permission creates stale-authority risk.
• Fiat, card, wallet, and stablecoin systems differ in reversibility, custody, fraud controls, and legal recourse.
• The sources provide design patterns rather than validated universal thresholds for safe delegation.
• Actual liability remains jurisdiction-, contract-, and fact-specific; technical controls supply evidence rather than a universal legal rule.

What Changed

• Added liability allocation as a design purpose for spending limits, warnings, confirmations, and records.
• Distinguished in-scope delegation, user safeguard bypass, and product deviation.

Related Concepts

• Agent Payment Infrastructure / 智能体支付基础设施 - payment and dispute layer implementing spending mandates.
• Agent Permission Boundaries - broader authority model within which monetary limits sit.
• Agent Identity And Authentication - attribution of the principal and acting agent.
• Verifiable Intent / 可验证意图 - semantic task and constraint record carried with payment authority.
• Agent Trust Calibration / 智能体信任校准 - basis for increasing or narrowing delegated limits over time.
• Stablecoins - settlement route with different reversibility and custody assumptions.
• Autonomous Agent Transaction Liability - responsibility framework that uses spend-control evidence.