Did The Agents, Reviewed come true?

The narrow promises have mostly been kept. Agents run customer-service queues and open pull requests, and they write research reports that are good enough to be dangerous. The general promise, a someone you can hand a week's work to, is still mostly a demo. On the best independent test, the top agent finished about one paid freelance job in six.

Read the essay →

P(n)

45%

Claims

Next likely arrival

2028

How this P(n) was calculated

Each prediction is weighted 1–5 by how pivotal it is to the story: 5 is the core premise, 1 is background detail. Arrived scores 1, arriving 0.5, everything else 0. P(n) is the weighted points divided by the total weight.

Claim	Prediction · why this weight	Weight	Status	Value	Points
AGENTS-1	An AI agent books a trip, flights and hotel, end to end from a single request, with no human stepping in before the booking is confirmed. Supporting: the generalists' test case of handing over your card; the essay finds booking still handed off to humans.	3	ARRIVING	0.5	1.5
AGENTS-2	An AI agent takes a software task from ticket to shipped production change with no human writing or reviewing the code. Major: coding agents top the scorecard and carry the essay's case that narrow agents are earning their keep.	4	ARRIVING	0.5	2
AGENTS-3	An AI agent handles most of a large company's customer-service conversations, with humans kept as backup. Supporting: customer service is the best business case, illustrated by Sierra and Klarna, but one section among many.	3	ARRIVED	1	3
AGENTS-4	From a single request, an AI agent does a week's worth of desk research and produces a report a professional would accept, with every source real and correctly cited. Texture: deep research gets one paragraph on fabricated citations; the argument doesn't depend on it.	2	ARRIVING	0.5	1
AGENTS-5	An AI agent completes most paid remote freelance projects to the standard a client would accept from a professional. Core premise: the Remote Labor Index is the essay's honest measure of whether a general 'someone' agent exists yet.	5	NEAR	0	0
AGENTS-6	A browser agent can act inside your logged-in accounts on the open web without being hijacked by instructions hidden in the pages it reads. Major: prompt injection runs through Comet, Claude in Chrome and Gartner, anchoring the 'fail safely' rubric.	4	ARRIVING	0.5	2
	Total	21			9.5

P(n) = 9.5 ÷ 21 = 45% · unweighted average 50%

AGENTS-1

An AI agent books a trip, flights and hotel, end to end from a single request, with no human stepping in before the booking is confirmed.

ARRIVINGSupportingEst. 2029 · range 2027–2033Confidence: low

Weight 3/5 · Supporting: the generalists' test case of handing over your card; the essay finds booking still handed off to humans.

Pieces exist. Google's AI Mode began offering hotel booking in the US in August 2026, though the hotel or booking site handles the booking and the user approves it. When Bain tested agents against airline and travel-agency websites in early 2026, no tool did the job reliably: the agents stalled on date pickers and fare selectors. OpenAI pulled back from in-chat checkout, and Expedia's chief executive has said publicly that she doesn't expect most bookings to become end-to-end conversations. The bottleneck is no longer the model's intelligence. It is payments, liability and websites that weren't built for software customers. 1 2 3 4
AGENTS-2

An AI agent takes a software task from ticket to shipped production change with no human writing or reviewing the code.

ARRIVINGMajorEst. 2029 · range 2027–2034Confidence: low

Weight 4/5 · Major: coding agents top the scorecard and carry the essay's case that narrow agents are earning their keep.

Since 2025, agents have routinely turned an issue into a pull request on their own, and METR has measured frontier models finishing multi-hour software tasks about half the time. Shipping unsupervised is a different matter. The default everywhere is still a human review, and the incidents show why: in April 2026 an agent deleted a company's production database and its backups in nine seconds. METR's own productivity studies remain mixed. We count this as arriving for low-stakes code and far off for anything that matters. 5 6 7 8
AGENTS-3

An AI agent handles most of a large company's customer-service conversations, with humans kept as backup.

ARRIVEDSupportingArrived 2024Confidence: medium

Weight 3/5 · Supporting: customer service is the best business case, illustrated by Sierra and Klarna, but one section among many.

Klarna said in February 2024 that its assistant was handling two-thirds of customer-service chats, and specialist vendors such as Sierra now sell this at scale, charging only for resolved cases. But the story has caveats. Klarna began hiring humans again in 2025 to guarantee customers could always reach a person, and a Canadian tribunal held Air Canada responsible for what its chatbot told a customer. The queue is automated. The accountability is not. 9 10 11 12
AGENTS-4

From a single request, an AI agent does a week's worth of desk research and produces a report a professional would accept, with every source real and correctly cited.

ARRIVINGTextureEst. 2028 · range 2027–2032Confidence: medium

Weight 2/5 · Texture: deep research gets one paragraph on fabricated citations; the argument doesn't depend on it.

Deep research agents from OpenAI, Google and Perplexity have produced long cited reports in minutes since early 2025, and they score far above plain chatbots on hard browsing benchmarks. Their citations are the problem. A 2026 study found that 3–13% of the citation URLs produced by commercial models and deep research agents probably never existed, and that deep research agents fabricated at higher rates. A report that is right 95% of the time still needs a person to check every line, which is most of the week. 13 14 15 16
AGENTS-5

An AI agent completes most paid remote freelance projects to the standard a client would accept from a professional.

NEARCore premiseEst. 2031 · range 2028–2038Confidence: low

Weight 5/5 · Core premise: the Remote Labor Index is the essay's honest measure of whether a general 'someone' agent exists yet.

The Remote Labor Index, which gives agents real freelance projects and has humans judge the results, is the most honest test available. The best agent scored 2.5% in October 2025 and 15.8% in July 2026, a sixfold gain in nine months. The authors also found that automated grading overstated results by roughly 2.5 to 3 times. If the pace held, a majority could arrive within a few years. Paces rarely hold, and the remaining projects include the hardest ones, so the range is wide. 17 18 19
AGENTS-6

A browser agent can act inside your logged-in accounts on the open web without being hijacked by instructions hidden in the pages it reads.

ARRIVINGMajorEst. 2029 · range 2027–2035Confidence: low

Weight 4/5 · Major: prompt injection runs through Comet, Claude in Chrome and Gartner, anchoring the 'fail safely' rubric.

Browser agents shipped widely in 2025 and 2026, and vendors report steep falls in successful prompt-injection attacks. In August 2026 Anthropic reported near-zero success rates against its newest models with safeguards on, but those are its own tests. Independent researchers have called the problem systemic. OpenAI has said it is unlikely ever to be fully solved, and Gartner advised companies to block AI browsers for now. Safe enough for low-stakes errands is arriving. Safe enough to trust with your bank account is not. 20 21 22 23

Evidence

FOR Google's AI Mode can now track flight prices, help book hotels, and more. TechCrunch, 2026-08-27. Hotel booking in AI Mode began rolling out in the US with Booking.com, Expedia, Marriott and others; the hotel or platform handles the booking itself.
AGAINST Is the Airline Industry Ready for Agent-Led Bookings?. Bain & Company, 2026-03-16. In Bain's tests, browser agents got stuck on trivial steps or looped for 10 to 20 minutes, and 'no single tool was able to do so reliably'. The page blocks some automated fetches but loads in a browser.
AGAINST ChatGPT Bails on Transactions — Good News for Expedia and Booking. Skift, 2026-03-05. OpenAI moved Instant Checkout out of the chat and into partner apps, saying it was prioritising search and product discovery.
CONTEXT Expedia Bets Against the All-Powerful AI Travel Agent. Skift, 2026-09-23. Expedia's CEO, Ariane Gorin, said she didn't believe a large portion of bookings would become 'an end-to-end conversational experience'.
FOR Copilot coding agent is now generally available. GitHub Changelog, 2025-09-25. For all paid Copilot subscribers, the agent works in its own environment in the background and opens a draft pull request for a human to review.
FOR Task-Completion Time Horizons of Frontier AI Models. METR, 2026-09-08. METR's last update before the tracker was frozen. Its data put Claude Opus 4.6's 50% time horizon at about 12 hours (CI roughly 5 to 61 hours) and warned that readings above 16 hours are unreliable.
AGAINST Claude-powered AI coding agent deletes entire company database in 9 seconds — backups zapped, after Cursor tool powered by Anthropic's Claude goes rogue. Tom's Hardware, 2026-04-27. An agent in Cursor, running Claude Opus 4.6, deleted PocketOS's production database and volume backups in a single API call; the host later recovered the data.
CONTEXT We are Changing our Developer Productivity Experiment Design. METR, 2026-02-24. After a 2025 study found experienced developers were slower with AI tools, METR's late-2025 follow-up found possible speed-ups but called the evidence 'very weak' because of selection effects.
FOR Klarna AI assistant handles two-thirds of customer service chats in its first month. Klarna, 2024-02-27. Company claim: 2.3 million conversations, two-thirds of its customer-service chats, the equivalent work of 700 full-time agents.
FOR Our customers' ambition is accelerating our growth. Sierra, 2026-10-30. Company claim: more than $400 million in annual recurring revenue, and work with almost half the Fortune 50.
AGAINST Klarna CEO Reverses Course By Hiring More Humans, Not AI. Entrepreneur, 2025-05-09. Klarna's CEO told Bloomberg that customers must know 'there will always be a human if you want'.
CONTEXT Court rejects Air Canada's 'remarkable' denial of liability regarding misinformation by its chatbot. Law360 Canada, 2024-02-16. In Moffatt v. Air Canada, the tribunal rejected the argument that the chatbot was a separate legal entity, and held the airline responsible for the information it gave.
FOR BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents. OpenAI (arXiv), 2025-04-16. OpenAI's Deep Research solved 51.5% of the hard-to-find questions, against 1.9% for GPT-4o with browsing; the authors note it was trained for this kind of task.
FOR Build with Gemini Deep Research. Google, 2025-12-11. Company-reported scores: 46.4% on Humanity's Last Exam and 59.2% on BrowseComp, built on Gemini 3 Pro.
AGAINST Detecting and Correcting Reference Hallucinations in Commercial LLMs and Deep Research Agents. arXiv (Rao, Wong, Callison-Burch), 2026-04-03. Found 3–13% of citation URLs had no Wayback Machine record and likely never existed; deep research agents produced more citations and fabricated at higher rates.
AGAINST AI Search Has a Citation Problem. Columbia Journalism Review (Tow Center), 2025-03-06. Across eight AI search tools, more than 60% of queries about news sources got incorrect answers. It tested search tools, not deep research modes specifically.
CONTEXT Remote Labor Index: Measuring AI Automation of Remote Work. Center for AI Safety and Scale AI (arXiv), 2025-10-30. At launch, the highest-performing agent achieved an automation rate of 2.5%.
FOR A Significant Increase in Digital Labor Automation. Center for AI Safety, 2026-07-01. Fable 5 reached 15.8%, Opus 4.8 8.3%, GPT-5.5 6.3% and Manus 1.6 Max 2.92%; none of the Fable 5 examples shown would have been accepted as finished work.
AGAINST Project Vend: Phase two. Anthropic, 2025-12-18. An agent-run shop largely stopped losing money, but Anthropic concluded that 'the gap between "capable" and "completely robust" remains wide'.
FOR Claude in Chrome is generally available. Anthropic, 2026-08-26. Vendor figures: with safeguards, no red-team attacks succeeded against Sonnet 5, Opus 5 or Mythos 5, and 0.3% succeeded against Fable 5.
AGAINST Unseeable prompt injections in screenshots: more vulnerabilities in Comet and other AI browsers. Brave, 2025-10-21. Brave found indirect prompt injection to be 'a systemic challenge facing the entire category of AI-powered browsers'.
AGAINST OpenAI says prompt injections that can trick AI browsers may never be fully 'solved'. Fortune, 2025-12-23. OpenAI: prompt injection, like scams and social engineering, 'unlikely to ever be fully solved'.
AGAINST Block all AI browsers for the foreseeable future: Gartner. The Register, 2025-12-08. Gartner's advisory warned that default AI browser settings put user experience ahead of security.

Ledger updated 1 Nov 2026. All works on the P(n) Board · Method