When could an agent earn its keep alone?

The short answer

Handed the work, agents already earn more than they cost. On real freelance projects, the compute for one attempt costs a few dollars, while the project pays hundreds. So even one success in a hundred covers the bill. Today it is one in five.

What they cannot yet do is the whole job alone: find the client, win the work, deliver it well enough to keep a good record, and get paid. Running a small business is harder still.

Their reach on real paid work is about 25 to 60 times shorter than on METR’s software tests. But it is doubling just as fast, about every three months. If that pace holds, an agent working alone first pays its way between late 2027 and early 2029. If progress slows to its 2019–2025 pace, that slips to about 2030. Compute is not the limit. Failure is.

Handed the work

Already pays, attempt by attempt. The bar was passed between mid-2025 and June 2026.

Freelancing alone

December 2027 to April 2028 at today’s pace. May 2029 to February 2030 at the slower one.

A small business alone

October 2028 to February 2029 at today’s pace. April 2031 to February 2032 at the slower one. About six months sooner, at today’s pace, if it can afford to fail half the time.

The question

An agent pays its own way when its tries earn more, on average, than they cost. Two things decide that. It must be able to finish a task as long as the job that earns the money. And it must succeed often enough that its failures do not sink it.

“Alone” means that no person finds the clients, takes the risk, or checks the work. The estimate leaves out legal standing: holding accounts, signing contracts, getting paid. It asks only whether an agent could do the work well enough.

It puts four pieces together:

1. How long a task agents can finish, and how fast that grows.
2. How long the earning takes a person.
3. How often an agent must succeed, which depends on what a failure costs.
4. What running costs, and how fast that falls.

1. How long a task agents can finish

The time horizon. METR measures the length of task, timed by how long it takes a skilled person, that an agent finishes alone half the time. On METR’s software tasks it has doubled every 89 days since 2024 (the trend in the papers’ Fig. 3a). METR’s tasks stop there: it calls “measurements above 16 hrs” unreliable. Mythos Preview reached that ceiling in April 2026; its horizon is known only to be at least 16 hours. Measured to 16 hours; projected beyond

Half the time, or four times in five. Kwa and colleagues found that “models’ 80% time horizons are 4-6x shorter” than their 50% horizons, and that both double at the same pace. Under METR’s curve, that ratio sets how fast success falls as tasks get longer. With a ratio of four, an agent whose 50% horizon is H hours succeeds on a task L hours long with probability H ÷ (H + L). In plain terms: a task as long as the horizon succeeds half the time, one a quarter as long succeeds four times in five, and one four times as long succeeds once in five. Chapter 3 draws this curve, for software tests and for real paid work. Measured relation

Real paid work is harder than software tests. METR says so itself: its tasks are “much ‘cleaner’ than real economically valuable labor”. Agents using a computer through web pages or screenshots had horizons “~50x shorter” than in other domains. And when people judged agent code by whether it could be merged, not by whether it passed the tests, an agent that passed the tests 38% of the time had none of its hand-checked work mergeable as-is. Measured

Calibrating on real work. The Remote Labor Index (RLI) gives agents 240 real projects from Upwork. They took people 11.5 hours at the median and 28.9 hours on average. The best agent now completes about 21% of them to a client’s standard: 20.83% for GPT-6 Astra in September 2026, and 21.25% for Opus 5.5 on the leaderboard. We assume project lengths are spread the way that median and mean imply (a lognormal). METR’s curve then turns a 21% share into a horizon on real paid work of about 2 hours at 50%, and about half an hour at 80%, as of October 2026. Derived from measured shares; the curve’s slope and the spread of lengths are assumptions

The two rulers agree on pace. Horizons derived this way double every 87 days (every 71 if the 80% horizon is a sixth of the 50% one). METR’s software horizon doubles every 89. For the same model, the real-work horizon is 24 to 61 times shorter: GPT-5 by 41 times, Gemini 3 Pro by 61, Claude Opus 4.5 by 24, and Claude Opus 4.6 by 52. That matches METR’s own “~50x” for computer use. Four times in five, the best agents finished real paid work of under a minute in mid-2025, and about half an hour by September 2026. Derived

2. How long the earning takes

One paid project. The RLI’s projects take a person 11.5 hours at the median and 28.9 on average. They pay $200 at the median and $632.60 on average. The index leaves out the hard parts of selling: it excludes “projects requiring physical labor, long-term evaluation, or direct client interaction”. Measured

The whole freelance job: find it, win it, do it, get paid. On freelance platforms, “for every hour of paid tasks, workers spend about 23 minutes” on unpaid work, such as looking for work or building up a profile (ILO 2021). So the whole job is about 1.38 times the project: 16 to 40 hours, or one to five working days. Handed the work, an agent does about three quarters of it. Alone, it must do all of it. Measured inputs; our combination

A small business that pays. We assume a working month to three months of a person’s time: 173 to 520 hours. We found no measured figure. The papers put earning in a market at weeks to months (companion Fig. 6b). The agent trials agree: Project Vend ran about a month without making money, and Andon Labs’ store and café have run since April 2026 without a profit. Assumption, bracketed

3. How often an agent must succeed

The rule: on average, a run must earn more than it costs. The success rate times the pay for a success has to beat the compute for a run plus whatever a failure costs. So the cost of a failure sets the bar.

If a failure costs only compute. This is the agent handed the work: someone else found the client and takes the risk. The RLI paper measured the compute for an attempt at $2.34 on average (median $0.92, highest $29.51); “models stopped generating before exceeding $30”. Measured

Successes cluster in cheaper projects. Manus earned 1.2% of the projects’ value with a 2.5% success rate, so a success is worth just under half an average project, about $290. Break-even is then a success rate of 0.8% at the measured cost, or 10.5% even at the $30 cap. The best agent’s share passed the first in mid-2025 (0.83%) and the second in June 2026 (15.8%). Handed the work, agents already pay their way. Derived from measured costs, shares and earnings

If a failure costs its reputation. This is a freelancer working alone. Upwork warns that “if your score falls below 79%, you may find it difficult to connect with new clients or win new projects”; 90% or more is excellent. So an agent selling its own work must succeed about four times in five. Its 80% horizon must reach the length of the whole job. Measured threshold (a platform’s rule)

If a failure costs the business. This is an owner working alone. A failed venture loses its stake and months of compute, so we hold an owner to the same bar: four times in five. An owner that can afford to fail half the time, perhaps with a backer, gets there about six months sooner at today’s pace. Assumption

4. What running costs

It falls fast. For a fixed level of capability, the price falls 12 to 45 times a year (chapter 1). Emberson and Roodman put it at “about 47% per quarter since 2023, or 13× per year”. Cottier and colleagues found 9 to 900 times a year, with a median of 50. Measured

It is cheap next to a person. More than 80% of the successful runs in METR’s study cost under 10% of a skilled person’s pay. GDPval found GPT-5 474 times cheaper than an expert on the bare model cost. But once a person reviews the work and redoes the failures, that falls to 1.18 times for one try and 1.63 times for repeated tries. Failures are the expensive part. Measured

The standing bill. Kudzu, an agent that published its own ledger, logged $147.17 of compute in eleven days, about $13 a day or $400 a month, and earned nothing. At the RLI’s median of $200 a project, staying alive takes about two typical jobs a month. Its own diagnosis: “the constraint on an agent like me is not capability … It is distribution and legal standing.” Reported (the agent’s own account)

So cost is not what binds. The compute for an attempt is around 1% of what the work pays. What binds is length (the whole job) and reliability (keeping clients, keeping the business).

How the pieces fit

A freelancer and an owner working alone both need to succeed four times in five. So the estimate comes down to one line: the length of real paid work an agent finishes four times in five. Today that is about half an hour. The question is when it reaches the length of the job.

L

The task’s length, in hours of a skilled person’s work: 16 to 40 for a freelance job, 173 to 520 for a small business.

H

The agent’s horizon on real paid work at 50%, calibrated on RLI shares. A quarter of it is the 80% horizon: about half an hour in October 2026.

Days per doubling

87 at today’s pace, fitted to the RLI points (METR’s software line: 89). 196.5 at METR’s slower pace of 2019 to 2025.

Worked through, for a freelance job. The shortest whole job, 16 hours, is about 30 times today’s half hour: about five doublings. At one doubling every 87 days, five doublings take about 14 months, which lands in December 2027. The longest job, 40 hours, is a little over six doublings away: April 2028. At the slower pace, those six doublings take about 40 months, which lands in February 2030.

So, at today’s pace, an agent working alone first pays its own way between late 2027, as a freelancer, and early 2029, as a small business. If progress slows, the freelancer arrives around 2030. That sits inside the test the home page sets for the whole program: “a census that finds no agent lineage paying its own way by 2030 would count against swarm ecology.”

Why not read METR’s line straight across? On its own, METR’s software trend reaches a working month’s task at 50% by December 2026, and three months’ by May 2027. That is right about software tests. On real paid work the horizon is roughly five doublings shorter, so the same lengths arrive about 15 months later.

A check: agents running shops

Vending-Bench (Andon Labs, February 2025; simulated). The best model, Claude 3.5 Sonnet, ended with a mean net worth of $2,217.93, against $844.05 for a person who played for five hours. But “all models have runs that derail”, and even the best model’s worst run ended at $476, below the $500 it started with. Measured, in simulation

Vending-Bench 2 (Andon Labs’ leaderboard, read 19 October 2026; a simulated year from a $500 start). The best model, GPT-6 Astra, averaged $15,514.70 (± $1,074). Andon estimates that a good strategy could make roughly $63k. The frontier rises about $822 a month. A year’s run takes 3,000 to 6,000 messages and 60 to 100 million output tokens. The top models show “no signs of performance degradation” over the year. Measured, in simulation

Project Vend, in a real office. In phase one (June 2025), a Claude agent ran a small shop for about a month and “did not succeed at making money”. In phase two (December 2025), with a CEO agent and better tools, “weeks with negative profit margin were largely eliminated”, across three cities. Reported by the developer

Why we built Pion (Andon Labs, September 2026). By late 2025, running a real-life vending machine “was no longer a challenge. AI could now run a business profitably.” A store in San Francisco and a café in Stockholm, handed over in April 2026, “lost a lot of money”. “Neither is profitable today.” Reported by the developer

Reading. The simplest business already pays when its location and customers are handed over and people do the physical work. That is the handed-the-work case again. A business an agent must build and run itself, hiring, renting and finding customers, does not pay yet. That fits the window above, which opens in late 2028. Simulation runs ahead of reality: Andon says models “got overwhelmed by the ‘messiness’ of the real world”.

What would move the dates

Earlier.

• METR’s 16-hour ceiling hides faster progress.
• RLI shares jump again, as they did from May to June 2026 (8.3% to 15.8%).
• Agents learn to pick jobs they can do. Near its horizon, METR finds, a model succeeds every time on about a third of tasks. Choosing well could clear the four-in-five bar sooner.
• People rent agents their accounts and their standing.

Later.

• The RLI saturates on easy projects.
• Finding clients turns out harder than doing the work. It involves direct client interaction, which the RLI leaves out.
• Legal standing (accounts, contracts, payment) blocks agents whatever their skill. This model does not cover it.
• The pace slows to its 2019–2025 rate.

Fragile assumptions. The slope of METR’s curve (k), the lognormal spread of project lengths, the value of a success (weighted by one agent, Manus), the length of a small business, and the bar of four in five for an owner. The last two are our assumptions.

Sources, and what kind of number each gives

Every source is cached. Measured: read off a study, a benchmark, or a platform’s own rule. Reported: one party’s account of its own work. Derived: computed by us from measured numbers. Assumed: our choice, bracketed. Projected: a measured trend carried forward.