The agent economy: when software buys from software

In May 2026, Robinhood launched agentic trading: brokerage accounts an AI agent can operate through MCP, the Model Context Protocol, an open standard for connecting agents to tools. Meanwhile, Claude Code can run scheduled agent sessions, so an agent can wake at a set time, do a job, and stop, with no human at the keyboard. Neither of these is speculative. They are shipping products, and together they sketch the outline of something larger: an economy where software transacts with software while humans set the policy.

That outline is what this essay is about. Part of it already exists. Part of it is prediction, and the predictions will be labeled as predictions.

The plumbing already exists

An agent economy needs three things: agents that can act without a person present, standard connections between agents and services, and services willing to accept a machine on the other end. All three now exist. Scheduled sessions cover the first. MCP covers the second: instead of every service writing a custom integration for every agent, one open protocol lets any compliant agent talk to any compliant tool. And Robinhood's agentic accounts are an early example of the third, a regulated business deciding that an agent, operating under rules its owner wrote, is an acceptable counterparty. The concrete day-to-day version of this is walked through in what agentic trading actually means.

None of this required artificial general intelligence. It required protocols, permissions, and a scheduler. Infrastructure is usually boring right up until it isn't.

Prediction one: discovery moves to answer engines

Here is the first prediction. Within a few years, most product discovery for software and services will be mediated by answer engines: assistants that fetch pages, read them as data, and return a synthesized comparison. Not a ranked list of links with ads on top. An actual answer.

This changes the physics of marketing. A human skims a landing page and absorbs the mood: the fonts, the confidence, the social proof. An agent does none of that. It reads the pricing table, the docs, the changelog, the refund policy, and whatever structured files you publish for machines. It can fetch a competitor's equivalents in the same second. Persuasion techniques tuned for human attention, like urgency banners, decoy pricing, and testimonial walls, mostly stop working, because the reader has no attention to capture and infinite patience for fine print.

Prediction two: purchasing authority gets delegated, with budgets

The second prediction follows from the first. Once an agent can discover and compare, the natural next step is letting it transact within limits: a monthly budget, a category allowlist, an approval threshold above which it must ask. There is a longer piece on agents with purchasing power, but the short version is that the delegation will look like corporate procurement, not a blank check. Software gets a budget line and standing instructions, and a human reviews the log.

To be clear about the present tense: in July 2026 most purchases still involve a human clicking a button, and most agent purchases are really humans approving what an agent recommended. The budgets-and-authority version is a forecast, not a description.

What changes for builders, and what changes for buyers

If those two predictions hold, the incentives for anyone selling software shift in a specific direction: machine-readable truth starts beating persuasion.

• Claims become testable, so honest claims become a ranking advantage. An agent evaluating a tool can read the methodology, check whether the numbers come with caveats, and notice when a page promises what no one can promise. A vendor whose claims survive verification gets recommended again. A vendor whose claims don't gets quietly dropped from comparisons its would-be customers never even see.
• Documentation becomes the storefront. The pages that matter most are the ones a machine can parse: clear pricing, explicit limitations, structured files like llms.txt, an FAQ that answers the hard questions instead of dodging them.
• Buyers comparison-shop without ad exposure. The agent never sees the retargeting campaign. It sees the product, the price, and the terms, lined up next to every alternative's product, price, and terms.

None of this makes markets kind. It makes them literal. A builder who has been rounding up will experience agent-mediated discovery as a hostile audit. A builder who has been publishing caveats all along will experience it as free distribution.

Where a one-time $29 tool fits

Full transparency about my own position here, because Coil (coil.trade) was built as a bet on exactly this shift. It is agent-native software: a long-only trading system you buy once for $29, no subscription, that runs inside your own AI agent against your own brokerage account. The scanner scores every S&P 500, Nasdaq-100, and macro-book name on leadership and entry quality, and the engine trades those published scores by rule. It buys leaders pulling back to real support, sizes by conviction, and goes to cash when nothing qualifies. Keys stay on your machine, and it ships disarmed; arming it is a deliberate human step. The research behind the ranking is laid out at /how-it-works, next to its caveats, because a product built for agent-mediated discovery has to assume its claims will be checked.

Two things worth saying plainly. First, a one-time price is itself a design choice for this economy: an agent comparing recurring costs against a flat $29 has a very easy row to fill in. Second, and more important, Coil is not income and no trading tool is. Markets can lose money, leveraged ETFs can lose it quickly, and nothing about agent-mediated buying changes that. If you want the mechanics of connecting an agent to a broker, the Claude plus Robinhood guide covers the setup end to end.

FAQ

What is the agent economy?

A shorthand for an economy where AI agents handle discovery, comparison, and some transactions on behalf of people, within limits their owners set. Pieces of it exist in July 2026, like Robinhood's agentic trading accounts. Broad delegated purchasing is still a prediction.

Do AI agents actually buy things today?

Mostly no. In most cases today, an agent researches and recommends, and a human approves the transaction. Agent-operated brokerage accounts are a real but narrow exception, and they run under rules the account owner writes.

Why would honesty become a ranking advantage?

Because agents verify instead of skim. An agent can read documentation, compare claims across vendors, and drop products whose claims do not hold up. Vendors that publish checkable, caveated claims survive that filtering better than vendors that rely on persuasion.