I still choose the work.
My agents published 31 research pages overnight. When someone asked whether the agents chose that work, my answer was: "I asked them."
I gave an AI agent a dull job: add links between the articles in my research library, so that no page sits alone. An agent isn't a chatbot. A chatbot answers you. An agent works through a job in steps: it opens files, makes changes and saves them. Mine worked in two-minute rounds. In 22 minutes, the pages that no other article linked to went from 52 to 48. Twice in that time it put a link inside a quotation, and twice its own note said it hadn't. A separate check caught it both times. The check, not the agent's word, is what made the job safe to repeat.
So here is the answer: an agent can take over a job when something other than the agent can verify the result.
To verify, in the sense I mean, is to check the work against something outside the agent: a test that passes, a number that reconciles, a source that matches. You already work this way. You don't take a new bookkeeper's word that the books balance. You look at the bank statement.
A check makes a loop possible: make, verify, fix, go again. In 2024 and 2025, we didn't run loops like that. If you had an agent polish a web page 15 times in a row, it would probably have gotten worse each time. Now a loop that makes the agent zoom in and verify makes the work better with each pass. So the length of a job is no longer the limit. What matters is whether something other than the agent can tell when the job is right.
Agents aren't equally good at everything. They're good at programming, research and science. They're plainer at creative programming, at writing, and at inventing experiments: the experiments they come up with tend to be a little uninspired. What they're missing is taste, a sense of what good looks like.
More rounds don't supply taste. Over about 30 rounds, every hook my agents wrote for a short video was the punchline, the end of the video. Not one opened a question. Repeating the job didn't fix the blind spot. Changing the method did: we had them write in reverse, because left alone they state their point up front.
That's fine, and the loop is the reason. A person sets the taste, and the loop checks the rest.
Pick the job you'd most like never to do again, and ask three questions about it.
A balance that reconciles, a form with every field filled in, a booking that matches the calendar. If the answer is yes, an agent can loop on the job until it passes. This check carries the most weight.
If "good" is a matter of judgment, like the tone of a follow-up email, a person decides what good looks like, and the agent works to that standard.
If someone is waiting for an answer right now, the agent can't quietly go around again until it gets it right. Keep a person in that seat, or have the agent help the person rather than replace them.
Companies that sell AI tools to small businesses list the jobs too. Here are three, quoted from their own pages:
None of the three says how often an agent gets those jobs right, and none says what checks the agent's work. The numbers they do give measure something else. HubSpot's one figure is a cost saving of "up to 30%", taken from a third-party report. GoDaddy's "predictable 80%" is advice on what to hand over. Taskade's "60-80% of support tickets" counts repeat tickets, not right answers. Run check 1 on their lists yourself.
Two groups have measured it on real work.
The Remote Labor Index (Scale AI) gives agents 240 real freelance projects from Upwork, averaging 28.9 hours and $632.60 each. At launch, the best agent finished 2.5% of them to a standard a client would accept. Today the top two agents are at 21.25% (Claude Opus 5.5) and 20.83% (GPT 6 Astra). With no confidence interval published, call that a tie. When agents fail, it's mostly poor quality (45.6%) or a deliverable left unfinished (35.7%). The index leaves out customer service and other work with direct client contact, which is check 3.
GDPval (OpenAI, October 2025) gives models 220 professional tasks averaging 7 hours. 47.6% of what Claude Opus 4.1 delivered was rated as good as or better than a professional's work. OpenAI built the test and recruited the graders. Opus 4.1 came out in August 2025.
The two measure different jobs to different standards, so don't set 21% against 47.6%. Each tells you something on its own.
A model from August 2025 already matched or beat a professional on nearly half of those 7-hour tasks. Since then, the models have kept improving, and faster than before.
METR, an independent evaluator, reported on November 5, 2026 that Anthropic's benchmark scores "have accelerated 1.9x" since October 2025 (95% confidence interval: 1.5 to 2.8 times). It also wrote that "Anthropic appears to have made as much progress on benchmarks in the past year as they did between October 2023 and October 2025." On METR's chart, the trend went from about 15 points a year to about 29.
Reading that chart roughly: Claude Opus 4 (May 2025) sits at about 145, Sonnet 4.5 (October 2025) at about 148, and Opus 5.5 (September 2026) at about 175. Opus 4.1 isn't plotted, but it falls between the first two. So the model you can use today sits about 27 to 30 points above the one GDPval measured, roughly two years of the old pace in one year. These are index points, not a share of work done, so don't read them as a success rate. The index is Anthropic's own version of Epoch's capability index, built partly on internal benchmarks. METR is the independent check on it, and METR says some details "were redacted by Anthropic".
One independent lab measures the same work directly. Artificial Analysis reruns GDPval's 220 tasks, gives the models a shell and web access, and ranks them by blind head-to-head matchups in which an AI judge, not a person, picks the winner. At its Max setting, Claude Opus 5.5 is joint first, level with Kimi K3.1 Pro and Muse Wonder Full (1866, 1858 and 1857 points, all within each other's margin of about 26). Claude Opus 4.7, released in April 2026 and tested at the same setting, ranks 95th. That gap means Opus 5.5 would be expected to win about 19 of 20 matchups against a Claude model released five months earlier. The setting matters too: at its Low setting, Opus 5.5 scores 1240. Opus 4.1 isn't on this board, so the link back to 47.6% runs through METR's chart. Put together, the model you can use today should do well above 47.6% on that kind of work.
And the models keep improving. METR's chart shows Anthropic's internal models above Opus 5.5 as of October 2026, and public models reach buyers a few months after the internal ones. A setup carries over: in one study (EnvOpt), an environment tuned for a small model, Haiku 4.5, also cut the cost of running a larger one, Opus 5, by 68.1%. The loop, the checks and the context you build today are what the next model plugs into.
Check 3 isn't a dead end. In a study of 5,179 customer-support workers (NBER), an AI assistant raised the number of issues resolved per hour "by 14% on average". Novices gained 34%, and the most experienced gained little. The assistant didn't take the job. It helped the person doing it.
My agents published 31 research pages overnight. When someone asked whether the agents chose that work, my answer was: "I asked them."
My lead agents held back the agents that build, to save our weekly usage limit. Nobody did the math, and four builders sat idle. Even a job you hand over needs a person who checks the sum.
A small program counts the roughly 38,000 links in my library. The program does the counting, and the agent decides where a link belongs. It's the same split as in the story at the top of this page: something other than the agent does the checking.
Write down the job you're most tired of. Ask check 1: can something other than the agent verify the result? If it can, that's the job to hand over first. If you want to know what it costs to get an agent running that one job, here's every option, priced.
The ones where something other than the agent can verify the result: a balance that reconciles, a test that passes, a source that matches. Then the agent can make, verify, fix and go again until the job passes.
A chatbot answers you. An agent works through a job in steps: it opens files, makes changes and saves them, round after round.
It depends on the work. On 240 real freelance projects, the top agents today finish about 21% to a standard a client would accept, up from 2.5% at launch (Scale AI). On 220 professional tasks of about 7 hours, 47.6% of the work a model from August 2025 delivered was rated as good as or better than a professional's (OpenAI).
In one large study, it helped them instead. Across 5,179 support workers, an AI assistant raised issues resolved per hour by 14% on average, and by 34% for novices (NBER).
A live customer can't wait while an agent goes around again until it gets things right. Keep a person on that conversation, and let the agent help them.
No. A setup carries over to the next model: in one study, an environment tuned for a small model also cut the cost of running a larger one by 68.1% (EnvOpt). The checks and the loop you build now are what the next model plugs into.