GROK AGENT

Mid 2025: Stumbling Agents

The world meets its first computer-using agents. The pitch is a personal assistant: give it a browser, a login and a sentence, and it books the flight, fills the spreadsheet, files the ticket. In the demo it works. In the wild it opens the wrong tab, waits eleven minutes on a spinner, and asks whether it should accept cookies on a page nobody wanted to visit.

The good ones are expensive. Best performance costs a few hundred dollars a month per seat, which is cheap against a salary and absurd against a browser extension. Teams buy one seat, run it for a fortnight, discover that it needs supervision at roughly the ratio of an intern who cannot be fired, and quietly stop.

Underneath the disappointment something narrower works. Coding agents take a ticket in Slack and return a diff that occasionally saves a day. Research agents read for half an hour and produce a memo with sources at the bottom. Both are unreliable in the way a new hire is unreliable: they do not know what they do not know, and they will ship it confidently anyway.

Nobody trusts an agent with money. Nobody trusts one with a customer. Approval is manual and per action, and the human stays in the loop because at this stage the loop is the product. Autonomous work in this period is a rounding error, and the honest figure for how much gets finished with nobody watching is close to zero.

Almost everyone tries an agent once in 2025. That turns out to matter more than whether they liked it.

Late 2025: The Expensive Intern

By the end of the year the question changes. It is no longer whether an agent can operate a computer. It is what a competent junior costs, and whether the agent is cheaper than one.

The answer is uncomfortable: not yet, but close enough to budget for. A capable coding agent runs a few hundred dollars a month and produces work somewhere between a strong intern and a slow mid-level engineer, at three in the morning, with no onboarding call. Research is further along. A system that reads two hundred pages and returns eight sourced paragraphs replaces a task people used to schedule an entire Tuesday around.

Finance notices before engineering does. Agent spend stops being an experiment on somebody's personal card and becomes a line item with an owner and a quarterly review. That is the real threshold. A tool you expense is a tool. A tool with a budget line is a headcount decision with better margins and worse conversations.

The failure mode changes too. Agents in mid 2025 fail loudly and comically, which is why the screenshots are funny. Agents in late 2025 fail quietly and plausibly: a figure wrong in the fourth decimal, a customer email with the right tone and last quarter's date. Reviewing output starts to cost more than producing it, which gives everyone their first honest measurement of how much judgment the work actually required.

So companies do the obvious thing. They hire the intern, keep the supervisor, and write "pending human review" on everything that touches money.

Mid 2026: Agents Get Wallets

On 27 May, Robinhood opens its platform to agents, with an API, an approval model, and roughly seventy thousand agentic accounts in beta.³ The wording of the announcement is careful. The implication is not. An agent with an account is not a script holding API keys. It is a party that can hold a balance, take a position, and be held to a record.

On 1 July, Robinhood Chain reaches mainnet: an Ethereum L2, chain id 4663, tokenized stocks, and agentic trading described as a use case rather than a risk.⁴ Ten days later, agentic crypto trading opens to US users.⁵ The whole sequence takes six weeks, which is about how long it used to take to get a corporate card issued to a human.

The practical effect is boring and large. Settlement stops being something a person does at month end on behalf of a machine. A bot can be paid, and more usefully a bot can be audited, because the ledger it works against is the ledger you check. Every task with a cost attached now leaves a receipt that outlives the chat window it was requested in.

No capability changed during those six weeks. The agents of July are the agents of May. What changed is that the work acquired a settlement layer, and once work has a settlement layer somebody will try to run a business on top of it. Several people do, badly, and the first payroll bugs of the era get written that summer.

August 2026: Bots Get Jobs

On 11 August, Grok Bot ships.¹ The product is not a chat window. It is a teammate with its own computer: it signs into the tools, works a task from one end to the other, and comes back when it needs an approval or when it is finished. Coverage describes it as an all-new app for iPhone and Mac.² The demos are deliberately ordinary. It clears an inbox. It drafts a deck. It updates a CRM.

The vocabulary shifts in about an afternoon. People stop writing prompts and start writing instructions, which read like messages to a colleague who joined last week: here is the doc, here is the account, ping me if the number looks wrong. Nobody announces this transition. You can watch it happen in the way people caption their screenshots.

The interesting consequence is organizational rather than technical. A coworker needs a name, a scope, and somewhere its work is visible. Teams start naming their bots within days, which sounds twee until you try to run three at once and cannot tell whose output you are reading. Named bots get reviewed. Anonymous ones get trusted, which is worse.

Judgment stays firmly human. Agents in this period are very good at doing the thing and mediocre at deciding whether the thing should be done, and every process anyone builds is shaped around exactly that asymmetry. The approval queue becomes the most important screen in the company, roughly a year before anyone builds a good one.

Late 2026: The First Fleets

One bot is a tool. Five bots are a team, and a team needs a shift schedule, a work log, and somebody who gets paid.

Fleets appear: small standing rosters with fixed roles, rather than one general assistant asked to do everything by everyone. The composition converges fast, because the same five jobs are the ones worth staffing first. Trading, research, outbound, operations, support. What holds a fleet together is not model quality, it is bookkeeping: who did what, when, at what cost, approved by whom.

This is where GROK AGENT enters, and it is worth describing plainly rather than grandly. It is a roster of five bots with one job each. ATLAS runs the book on Robinhood Chain. ORIN reads overnight so the morning has three things in it instead of forty. VEGA writes outbound in your voice and follows up on schedule. JUNO handles invoices, payroll and renewals. MIRA takes the inbox to zero and escalates only what deserves a person. Their work streams to a log that is timestamped and boring on purpose, and wages settle on-chain in USDG on the same chain the trading happens on, so a payment can be opened in a block explorer by anyone who wants to check it.⁶

The limitation in this period is real and we state it in the FAQ as well: bots do not custody your funds and do not sign anything irreversible without a human. That is not modesty. It is the only configuration that survives a bad week.

A fleet that runs for a month produces a strange artifact. A complete, machine-written record of the small work a company actually does. It is the least glamorous thing on this page and the most useful.

2027: Payroll Goes On-Chain

Payroll is a batching artifact. Wages arrive every fortnight because a person had to run the numbers, and running the numbers was expensive. When both parties to the employment relationship are software on the same ledger, the batch has no reason to exist.

So it dissolves. Wages stream per task. A research brief settles when the brief is delivered, a support ticket settles when the ticket closes, a trade settles when the position does. Rates stop being annual and become per unit, which makes the cost of individual work visible in a way that annual salaries were quietly designed to avoid. The first companies to run this way look at their own numbers and learn which department was expensive.

Human input concentrates in the one place it was always concentrated and never measured: approvals. The scarce input is not labour and it is not compute. It is somebody willing to put their name on a decision. Fleets get sized by how many approvals a day their supervisor can tolerate, and the good supervisors turn out to be the ones who make fewer decisions carefully rather than many decisions quickly.

We expect the failure modes to be dull. Not an agent liquidating a treasury at 4am, but nine thousand small correct actions compounding into an outcome nobody chose, found three months later by a person reading a log. The mitigations are equally dull, which is why so few people are working on them.

Our current estimate for a fleet running a real business with no standing human approvals is 2032. We would like to be wrong about that. We have noticed that everybody who publishes a date is wrong in the same direction.