1948, and the case for autonomous agents

This edition explores the rise of autonomous agents, and how the same shift that took elevators from human-operated machines to trusted systems is now arriving in software, finance and the wider economy.

In 1948, Otis began putting prerecorded voices inside some elevators. The cars still had attendants, but Elevoice welcomed passengers and sometimes asked them to let the doors close, partly preparing riders for a future without attendants.

Two years later, an office building in Dallas installed the first Otis high-speed elevator system designed to run without an operator. By the mid-1950s, operatorless elevators had gone from an uncomfortable novelty to something commercial buildings actively wanted.

Trusting a machine suspended hundreds of feet above the ground to accelerate, slow down and open its doors once felt strange. Over time, the system proved reliable enough that handing over control stopped feeling like handing over control.

AI has entered a similar phase. Most people still consult it: ask a question, receive an answer, decide what to do, then do it. Agents let the instruction set off the work, with software deciding the steps and continuing until the job is finished.

MExecution changes the role of the software. An agent with authority over a process keeps working after the person who gave the instruction has moved on.

Citi introduced Arc as a platform for building agents across the bank, covering research, synthesis, preparation and execution. More than 80% of the 180,000 employees with access to Citi's existing AI tools use them regularly, and Arc moves that use into workflows where agents act under monitoring and audit.

J.P. Morgan described the same shift in corporate treasury. At 6:47 a.m. in London, an agent has detected a supplier's change in invoicing currency, calculated a potential $2.3 million FX mismatch over the quarter, proposed a hedge and sourced three counterparty quotes. The treasurer arrives to find the exception waiting for approval, with evidence and an audit trail. Elsewhere, a digital twin of the balance sheet has run 4,000 scenarios overnight.

Human attention has always imposed a minimum economic size on a transaction. Few people spend ten minutes deciding whether to pay four cents or six cents for information, or compare five providers to save a fraction of a cent. The effort costs more than the thing being bought.

Software doesn't have that constraint.

Mastercard is building Agent Pay for Machines around large numbers of small, continuous transactions, including payments worth fractions of a cent. In Mastercard's example, an entrepreneur asks an agent to launch a flower shop online, and the agent can buy a domain, hosting, images and checkout infrastructure from different providers.

In an Unhashed conversation, Sri Misra discussed this with Nikhil Chandhok, Circle's chief product and technology officer. An agent handling a complicated job, Chandhok explained, can break it into smaller tasks, find specialist agents and pay them, and those agents can repeat the process further down the chain.

Chandhok remembers dial-up internet in India when graphical browsers were barely usable. Early Google was mainly useful to him for finding research papers, Amazon sold books, and eBay often looked like a market for junk. It took years before the economic structure around the web became visible.

Agents are easy to underestimate when the examples are restaurant reservations, travel itineraries and spreadsheet work. Chandhok's more interesting picture is a network containing billions of agents that perform work, develop reputations, hold wallets, receive money and pay one another. He described them on Unhashed as "economic actors."

An agent in Singapore can buy a service from one in India, which can pay another in the United States. The person who started the task does not need to know where any of them are hosted.

The internet made information global long before money moved with the same ease.

Early app ecosystems had poor support for tiny payments, while advertising already offered a way to monetize attention. That helped advertising become the business model underneath a huge part of the web and mobile economy.

As he put it, "how money works is very critical to how an ecosystem evolves."

If one agent can pay another $0.003 for a query, $0.10 for inference or $2 for a specialized dataset, work once too small for a subscription acquires a price. A research task can assemble ten services for a few cents each and discard nine outputs. Software can buy information for seconds, computation for milliseconds or a capability for one job.

What machines can afford to buy from one another will shape the economics of agents.

> in practice

At aarnâ, we build with agents and we run with them. Agents reviewed the ICT contract suite across four independent passes before it reached mainnet.

Three agents now sit under âTARS. The support agent answers investor and partner questions using aarnâ's own documentation. The operations and security agent sits between an intended action and a signed transaction: it simulates each action, checks it against policy, verifies that invariants hold, and confirms reconciliation before anything is proposed for signing, with every action written to an append-only trail. Risk, still in build, will read the underlying portfolio between reporting dates, so nothing reaches the investment and risk teams late. Credit decisions over the underlying will continue to sit with the funds' authorised, SEBI-regulated managers.

More at aarna.ai

An elevator lives inside a shaft. Steel rails determine where it can go, brakes constrain its movement and the possible failures are bounded by the physical system around it.

An economic agent faces volatile pricing, counterparty credit risk, adversarial inputs and other agents pursuing their own goals.

That means identity, spending limits, approved counterparties, velocity controls, audit trails, policy boundaries and clear escalation rules. An agent can begin with authority to spend $50 without asking, then $5,000 inside a defined procurement policy, and eventually execute routine treasury or portfolio actions within a tightly prescribed mandate. Each increase in authority depends on predictable behavior and a clearer answer to what happens when something goes wrong.

Trust accumulates through ordinary decisions that occur exactly as expected.

Otis had a prerecorded voice because passengers once needed reassurance that an elevator could operate safely without a person beside the controls. The voice eventually disappeared too.

Agent autonomy follows the same direction. More economic activity moves beneath the level of individual human approval, and over time the remarkable part will be how unremarkable that feels.

This issue grew out of a conversation with Nikhil Chandhok, Circle’s head of product and technology, on stablecoins, tokenised assets, liquidity, credit and the financial infrastructure being built around them. The full episode is on Unhashed.

reflections-

Gain an edge in DeFi alpha with aarnâ’s AI-driven insights and DeFi vaults. Try the dApp now.