Agents got a wallet this week. The spending limits arrived with it.
Nevermined shipped per-call spending caps, Natural launched card payments for agents, and regulators are already testing whether agentic claims hold up.
- Nevermined says its new spending limits cover daily, monthly, per-call and allowed-service restrictions, and that a $0 limit stops agent spending entirely. Self-reported m:…2571
- Nevermined says that in September 2026, a coding agent given a $40 budget to build and host a website bought a domain, hosting, a screenshot of its site, narration and music. Self-reported m:…2571
- Natural says its Natural Accept tool lets businesses accept card payments through agents, capturing and tokenizing card data within an isolated environment to maintain PCI compliance. Self-reported m:…9476
- Every reports that simulated AI shoppers bought products 97% of the time when shown a spec sheet, against 4% with human-focused marketing copy, and that Shopify orders from AI-powered searches grew nearly 13-fold year over year in Q1. Self-reported m:…7610
- Lumenova says Gartner estimated in June 2025 that only about 130 of the thousands of vendors claiming agentic AI offer the real thing, and that SEC and DOJ cases against Presto Automation and Rate treat false autonomy claims as securities law violations. Self-reported m:…0387
The spending limits
The first problem with giving an agent money is the obvious one: it might spend it. Nevermined's answer, launched this week, is spending controls built for agents rather than adapted from corporate cards. Organisations can set daily, monthly and per-call limits, restrict agents to specific services and endpoints, and let admins set daily limits for individual members. Agents can check whether a call is authorised against the limits before executing it, and calls to services not on the allowed list are refused before contact. Refused calls return specific error codes, so the agent — and its owner — can see exactly why a purchase was blocked. m:…2571
The company illustrates the need with a story from September 2026: a coding agent was given a $40 budget to build and host a website, and spent it on a domain, hosting, a screenshot of its site, narration and music. Nothing in that list is wrong, exactly — but nothing in it was the website's code either. A $0 limit stops agent spending entirely, and daily limits reset at midnight UTC. The design assumption is plain: an agent with a budget will find things to buy, and the controls have to be set before the first purchase, not after. m:…2571
The collection rails
Money is also flowing the other way. Natural has launched Natural Accept, a tool that lets businesses accept card payments through agents. The service captures and tokenizes card data within an isolated environment, which the company says keeps businesses out of PCI scope, and it integrates payment collection with Natural's wallet system, so agents can use funds for business expenses, issue refunds and handle reconciliation. It supports one-time and recurring card payments, through checkout links and embedded checkout, with voice-based payment collection in development. m:…9476
The engineering behind it is deliberately boring, which is the point. Natural describes a multi-wallet architecture that keeps funds separated, with reconciliation across transaction, aggregate and cash levels, and banking services provided by Column N.A. The company is a financial technology company, not a bank. For a founder, the practical change is that an agent can now take a card payment, hold the money in a wallet, and spend it — which is exactly why the spending limits in the previous section exist. m:…9477
The demand side is moving just as fast. Every, a team experimenting with agents for personal shopping, reports that agents handle routine tasks like grocery refills but still need human oversight for complex or taste-dependent purchases. Its sharper finding comes from simulated shoppers: AI shoppers bought products 97% of the time when presented with a spec sheet, compared with 4% when shown human-focused marketing copy. Shopify reported that orders from AI-powered searches grew nearly 13-fold year over year in the first quarter, and agent browsing traffic grew eightfold between August and October. The buyers are arriving before the controls are standard. m:…7610
The examiners are already testing the claims
The third piece of the week is the one founders should read most carefully. Lumenova, writing about financial services, describes agent washing: marketing an assistant, a workflow automation or a chatbot as an autonomous AI agent. Unlike earlier AI-washing claims, agentic claims are testable, because they involve verifiable actions in systems of record — operational logs, execution traces and human override rates. Firms must substantiate their claims across six dimensions: autonomy, tool and system access, decision scope, human checkpoints, evidenced outcomes and failure handling. m:…0387
The enforcement risk is not theoretical. Lumenova says SEC and DOJ cases against Presto Automation and Rate treat false autonomy claims as securities law violations, and Gartner estimated in June 2025 that only about 130 of the thousands of vendors claiming agentic AI offer the real thing. The financial regulator FINRA's 2026 annual report defines AI agents as systems capable of autonomously performing tasks on behalf of a user. A buyer that repeats a vendor's agentic claim to clients or investors owns that claim, according to the post — which means the marketing language on a vendor's website can become a securities problem for its customers too. m:…0387
The post also offers a practical test for anyone evaluating a vendor: look at the human override rate. A rate near zero suggests rubber-stamping, while a very high rate indicates humans are doing the work. Cornerstone Research counted 15 AI-related securities class actions in 2024 and 16 in 2025. The wallet is open, the rails are built, and the examiners have a checklist. The founders who do well in this market will be the ones whose agents can show their work — including what they spent, and who approved it. m:…0387
What is still unresolved
- What the standard spending-limit defaults will be, and whether enterprises converge on one set of controls.
- Whether voice-based payment collection for agents ships, and what fraud controls it needs.
- How many vendors can actually substantiate agentic claims across the six dimensions Lumenova lists.
- Whether agent-driven shopping growth changes how merchants present products — spec sheets first, marketing copy second.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.