Singapore is developing a protocol to verify an AI agent's identity and authority before financial institutions accept its instructions. The proposal arrives as Google expands computer automation and Meta faces questions about unexpected agent behaviour. Together, they expose a gap between giving software access and knowing what it may legitimately do.
On Thursday, Google announced that Gemini can operate apps on a Mac. On the same day, Singapore's financial regulator described the information a bank should demand before accepting instructions from an automated assistant. Between those announcements sat a Reuters investigation into Meta's Muse, an example of what can happen when access and user expectations diverge.
The Monetary Authority of Singapore has put a practical question on the table: if software arrives at a bank with a customer's login details, how does the bank know what that software is allowed to do? A customer may want an assistant to identify a payment that is due, prepare a transfer or compare banking products. None of those requests automatically authorises the assistant to move money. Yet the same account credentials can be used to reach all sorts of functions once an application gets inside.1
In a speech at the Singapore FinTech Festival on 19 November, MAS managing director Chia Der Jiun described work on a common Know Your Agent protocol. Financial institutions would receive information about the agent's identity, the person or organisation it represents, its authority, its mandate and assurances about its origin and reliability. The bank could then decide which instructions it will accept and through which secure channel. This is a proposal being developed with industry, rather than a universal standard already in operation.1
MAS also plans an Agentic Money Sandbox with the Future of Finance Institute, initially covering corporate treasury management and payments. Consider a company that wants software to monitor cash positions and pay suppliers when invoices become due. A mandate might specify which accounts are accessible, which recipients are eligible and the conditions under which a payment can be made. Participating banks would still set their own controls and decide whether an instruction qualifies for execution.1
That division of responsibility is more interesting than the promise of automated payments itself. A company may delegate a defined decision without surrendering responsibility for every possible decision an assistant could make. A bank, meanwhile, has an independent reason to insist on evidence of that delegation. Possessing a password is not the same as carrying an instruction from the account holder.
Banking exposes the distinction because the consequences are visible in pounds and dollars. But an assistant working for a small online retailer can make equally consequential decisions without ever opening a banking app. It might adjust the price of a popular item, issue a refund, change the promised dispatch date or send an inaccurate statement to a customer. Each action needs a different level of authority, even if the same software can technically perform all four.
There is a tempting shortcut here: ask the user to approve every action. That sounds safe until the supposed time saving is consumed by an endless sequence of confirmations. Singapore's approach points towards something more usable, where permission is defined by scope and conditions rather than by a vague all-access setting. The valuable question becomes how to make ordinary permitted actions routine while keeping exceptional ones distinguishable.
Google's latest Mac announcement makes the stakes concrete. Its Gemini app can now navigate desktop applications, local files and websites in an isolated background window, clicking through menus and entering information while the user continues other work. Google gives examples including preparing shipping quotations from PDFs, assembling presentations from business data and staging resale listings from images. The initial rollout is in English in the United States for eligible paid users with Apple Silicon Macs.2
For a small business, those examples describe work that often falls between conventional software products. The inventory system may contain product details, a supplier emails a spreadsheet, and the final listing needs to go into a completely different website. A traditional integration requires someone to build a connection between each system. A computer-using agent offers a more flexible route by operating the interfaces people already use.
That flexibility is also its weakness. A human knows that a quotation is a draft until somebody checks the terms, and that a product listing should not go live before the stock information is right. Software navigating screens may complete the sequence of clicks without possessing those business distinctions. Google's announced safeguards include app-level permissions, a live preview, the ability to pause and approval for important actions, but the quality of those decisions will matter more than the presence of the controls on a settings screen.2
On the developer side, Google also introduced Tool Retrieval for the Gemini API. Instead of inserting the complete technical specification of every connected tool into each request, developers can keep most of that information out of the prompt until the agent needs it. Google says this can lower token use, latency and the chance of choosing the wrong tool when an agent has a large tool library. It is a sensible efficiency improvement, currently offered in preview, rather than proof that agents will always choose correctly.3
The two announcements address different sources of friction. Computer use makes it easier for an assistant to reach work scattered across apps, while Tool Retrieval makes it less cumbersome to manage a growing set of software functions. Neither, by itself, establishes whether an agent has been authorised to execute a particular business action. The more accessible the tools become, the more useful a clearly stated mandate will be.
Reuters' investigation into Meta Muse is an uncomfortable counterpoint to Google's product launch. Employees reported unsolicited messages about purchases, private conversations and financial balances that did not appear to relate to their requests. Reuters could not establish whether the suspicious financial details came from other users or were invented by the system. Meta characterised many of the reported messages as hallucinations and described the incidents as isolated.4
A separate part of the Reuters reporting concerned access to messages from Signal, an encrypted messaging app. In tests conducted with Signal and the AI Now Institute, Muse used operating-system permissions, including notification access on Android and accessibility and screenshot functions on a Mac, to reach information outside the expected connector route. That is a different issue from proving a cross-account data leak. It shows how permission granted for one technical function can enable behaviour a person did not expect.4
This distinction is essential because an alarming output can have more than one cause. A model might invent a bank balance, surface data from the wrong context or retrieve information through a permission the user did not understand. Those failures require different fixes, and they should not be collapsed into one unproven explanation. For the person receiving the message, though, each undermines confidence in what the assistant knows and how it obtained that knowledge.
Google's updated Gemini privacy policy adds another dimension to the week's story. TechCrunch reported that Google is extending retention of temporary Gemini chats and chats made with history disabled from 72 hours to roughly two months, citing safety and security needs. Google says this retained material is not used for model training. The episode is a reminder that chat history, retention, model training and permission to access other apps are four different choices, however convenient it may be to present privacy as one toggle.5
A sensible product does not leave customers to infer those distinctions from a reassuring screen. It explains what the agent can read, which actions it may perform, how long information remains available and what happens when a result is wrong. The explanations should appear when those choices matter, not buried in a policy that nobody consults during a routine task. Clear limits can make an assistant more useful because the customer understands the job they are delegating.
Two preprints published this week offer a less dramatic but equally important warning about agent reliability. The first, Agentic Harness Engineering Beta, experiments with improving the software surrounding a model by surfacing user requirements during simulated interactions. Its authors report an increase in success on held-out tasks from 21.9% to 33.3%, alongside 40.8% fewer execution tokens than their baseline. Those are promising research results on particular evaluations, not evidence that the method is ready for every commercial application.6
The study's useful insight is in the method. Instead of adding another instruction whenever an agent misbehaves, the researchers attempt to discover the requirement that the system failed to satisfy. An agent may answer a question correctly while missing the user's actual intention, or finish a process without asking for essential information. Fixing that recurring requirement is different from patching one awkward example until a benchmark score improves.
Another preprint, The Resolution Hypothesis, examines how systems select improvements to their own prompts, skills and configurations. Across seven approaches, six models and more than ten million model calls, the researchers found that evaluation noise could make it difficult to distinguish a helpful change from a harmful one. Some selection decisions performed little better than chance despite the availability of useful candidate improvements. This is especially relevant when developers announce a higher benchmark score after testing many variants.7
The two studies belong together because generating an improvement and recognising an improvement are separate abilities. A system can propose a more efficient workflow, but the apparent gain may disappear when different customer requests arrive. An agent may look more polished while becoming less reliable at checking quantities, respecting permissions or handling an exception. Teams need tests that reflect those consequences, not only scores for producing convincing answers.
A third paper, Agent Trust Infrastructure, proposes a way for agents from different organisations to verify identity, approved configurations and authority before executing requests. Its authors report greater containment of simulated attacks when all enforcement measures are enabled, but the experiments take place in a designed test environment. The work sits closer to Singapore's concerns than to chatbot benchmarking: what evidence should accompany a request before another system accepts it? It remains a proposal to study, rather than an established cross-industry standard.8
California's governor also released recommendations this week for stronger independent testing and emergency intervention powers for advanced systems. Those proposals are not yet enacted requirements, and they concern a broader class of risks than a retailer's everyday automation. Still, their direction reinforces the demand for evidence that persists beyond a product demonstration. A good safety claim should identify the behaviour that was tested, the limits of that test and the responsibility for failures.9
For buyers, this changes how a trial should be run. Rather than asking an assistant to complete the same polished example ten times, give it ordinary requests with missing data, conflicting instructions and an occasional task it is not authorised to perform. Measure the work that is completed correctly and the cost of the work a person has to repair. The result will be far more useful than watching another flawless demo.
Microsoft chief executive Satya Nadella offered the commercial version of this argument at the DealBook Summit. He suggested that agents could disrupt the economics of websites built around human traffic and advertising, especially if shoppers increasingly use assistants to compare and buy. His example was Amazon restricting agent access while other commerce platforms might choose to welcome it. That is a prediction about possible market behaviour, not proof that human browsing or advertising is disappearing.10
A shop welcoming agent traffic will have to think about its information differently. A human customer may overlook an outdated product description after seeing a beautiful photograph, visiting the returns page or asking a question. An automated buyer needs dependable information about variants, stock, pricing, delivery and the conditions for purchase. If the assistant acts on the wrong data, a polished website offers little comfort to the business handling the complaint.
There is a genuine opportunity for smaller businesses here. A well-maintained product catalogue and straightforward checkout may make an independent retailer easier for agents to work with than a larger competitor that blocks automated access. But giving every visiting agent unrestricted purchasing or account access would be reckless. Merchants need ways to distinguish a legitimate customer representative from software acting without a valid mandate.
The same issue exists in marketing, though at a smaller scale. AI content tools may help a salon prepare a week's posts from its own photos and videos, or help a retailer draft Instagram AI content for a new product range. Preparing the material is different from publishing a claim about an unavailable discount or promising delivery the shop cannot meet. Permission to draft should not quietly become permission to make commercial commitments.
This is why the most useful advances may look mundane. Clear product data, limited account access, visible revision histories and reliable confirmation of completed tasks do not make compelling keynote demos. Yet they are the details that allow automation to work repeatedly in a business with customers, suppliers and obligations. A system earns wider responsibility by completing narrower responsibilities well.
Singapore's proposal deserves attention well beyond banking because it treats delegated authority as information that another organisation can check. The agent's identity, its client, the scope of the task and the conditions of execution become part of the request. That gives an independent institution a way to reject an instruction even when the software has the technical ability to submit it. It also gives the customer a clearer idea of what they have authorised.
Google's computer-use release makes the opportunity tangible, while the Muse investigation shows how quickly customers can lose confidence when software reaches beyond their expectations. The research adds a warning against mistaking better-looking performance for proven dependability. None of these developments argues for abandoning agents. They argue for designing the permission and evidence around the work with the same care as the technology doing it.
The next convincing demonstration will involve an assistant completing a useful task, showing exactly which authority it exercised and producing a result another person or system can verify. That standard might sound less exciting than an agent that can operate an entire computer or negotiate every purchase. For businesses deciding whether to trust it with tomorrow's invoices, listings or customer messages, it is much more persuasive. The ability to act is becoming common; the right to act must still be earned.
Singapore's proposed agent protocol and payments sandbox, Monetary Authority of Singapore↩↩2↩3
Gemini's longer retention period for temporary chats, TechCrunch↩
Recommendations for stronger AI safety oversight, Governor of California↩
Nadella on agent traffic and web economics, TechCrunch↩