Every AI Agent Now Comes With an Allowance

Three announcements landed on the same day this week, from three companies that have nothing to do with each other, and they all shipped the same feature. Anthropic's new Claude Duet must ask before it sends, pays, signs or deletes. Walmart's new Agent Profiles give every shopping agent its own spending limit. isolved's Agent View gives every HR agent a manager. The headline is not any single product. It is that the approval gate and the spending cap have quietly become standard equipment for AI agents, at every layer of the stack.

Anthropic: an assistant that drafts, then waits

On 21 October Anthropic opened a waitlisted research preview of Claude Duet, aimed at "professionals and small business owners managing crowded workloads." It is US only for now, on Pro and Max plans, and its background work counts against your plan limits. TestingCatalog and MacRumors have the fullest write-ups.

What separates Duet from a chat window is that it runs on its own. It prepares a morning brief from your connected apps, re-checks every few hours, and messages you when it judges something needs attention. It has its own virtual machine and browser, so it keeps working while your laptop is closed. It carries over your memory, skills and connected apps, and Anthropic says it can reach more than 6,000 connectors and plugins, with Gmail, Slack, QuickBooks, Stripe, HubSpot, Canva and Shopify among the named examples.

Then comes the part that matters. Duet reads, drafts and prepares by itself. But before it sends a message, pays for something, signs anything or deletes anything, it has to show you the proposed action and wait. Securities trading and account transfers are off the table entirely. Business purchases run through Stripe Link or Shop Pay, and Shopify said this is the first time it has opened Shop Pay inside an agent to stores that are not on Shopify.

Anthropic is candid about the rough edges. Duet may misread a thread or rank urgency badly, and you cannot yet watch its browser. MacRumors frames it as Anthropic's answer to Meta's Muse and OpenAI's Dots.

Walmart: the agent as an authorized user

The same day, Walmart CEO John Furner announced the Walmart Agent Gateway, an MCP server that lets approved agents search inventory and buy on a customer's behalf, as TechCrunch reported. Alongside it come Agent Profiles, rolling out next month.

Walmart's own analogy is the best one: an agent is treated like an authorized user on your credit card. Each agent gets a separate profile inside your account. Its purchases show up apart from your own. You choose whether every transaction needs your approval or whether the agent can spend freely under a limit you set. Unite.AI reported example caps of $50, $100 and $200.

The terms are strict. No actions beyond what the shopper asked for. No access to payment information. Only the data the task needs. Agents that break the rules get blocked, and Walmart's returns policy still applies if an agent buys the wrong thing. Amazon, meanwhile, has been blocking third-party agents, including Meta's Muse. The two largest US retailers have now picked opposite sides.

isolved: the agent gets a manager

Also on 21 October, HR software vendor isolved launched Agent View, which puts AI agents on the org chart next to people. Admins can see each agent's position, its manager, its role, its defined permissions and its performance record, starting with isolved's own payroll agent and the third-party agents it works with. The company cites its own survey finding that 82% of organisations now use AI in HR processes, up from 69% in 2025, though that figure is self-reported and I have not seen it verified.

The idea sounds almost quaint, but it answers the first question every business owner asks about an agent: who is responsible when it gets something wrong? On the org chart, the answer is a name.

The military is saying the same thing in policy language

This is not only a consumer and enterprise pattern. At AUSA, Lt. Gen. Christopher Eubank described how US Army Cyber Command runs its agents today, as Defense News reported. All work is "safeguarded human-in-the-loop." Personnel watch agents perform routine tasks, and a governance board inside the command decides, per task, whether an agent may continue without intervention or work unsupervised. Every agent checks in with its human team daily, and each is assigned the same work role as a human and trained like one. Eubank claimed one agent finished in two days what had taken an analytics support officer six months. That is his claim, not an audited result.

In Congress, the Senate Armed Services Committee has added provisions to the pending annual defense bill that would require human oversight of AI-controlled weapons, ban AI for nuclear launch or detonation, bar AI from domestic surveillance, and require the Pentagon to keep data on decisions made by or with the help of AI, according to a WUNC report republished by WHRO. Keeping a record of every AI-assisted decision is an audit log by another name.

Autonomy is no longer something an agent has. It is something an agent earns, one task at a time, from a human who can take it back.

Why this matters

Think about a teenager's first year of driving. You do not hand over the keys and the credit card on the same afternoon. You ride along, then you allow short trips, then you set a fuel budget, and you keep the right to take the keys back. Every one of this week's announcements is a rung on that ladder. Duet's four gates are the ride-along phase. Walmart's spending cap is the fuel budget. isolved's manager field is the parent whose name is on the insurance. The Army's governance board is the licensing authority that decides when supervision can stop.

For anyone building or buying agents, the practical lesson is that these controls are now expected, not optional. Not long ago, an agent with an approval step looked cautious. This week a frontier lab, the largest US retailer and a payroll vendor all shipped one. Buyers will soon compare agents the way they compare credit cards: what can it do alone, what needs my sign-off, what is the limit, and how fast can I cut it off.

The second lesson is about records. Walmart's agent terms and the Senate's decision-data requirement both assume someone can show, after the fact, what an agent saw, what it proposed, who approved it and what happened next. If you run a store, a practice or a team that is about to let agents act for you, the question to ask any vendor is simple: show me the ledger.