Back to the trend of the day: Some ChatGPT Pro users can now make always-on helpers on their phones
Poly Rating · from the trend of the day

What Can AI Agents Do? Top 10 Jobs in 2026

METR estimated Claude Opus 4.5 could finish tasks worth nearly 5 hours of human work, yet even the longest 0.1% of Claude Code turns ran only about 42 minutes: people leave agents alone far less than the agents can handle.

  • No. 1: AI coding agents (bug fixes, tests, features)
  • 10 places

01Final list

The leading AIs each built their own list of ten independently. Here is one result that is easy to check: every place has a reason, a source, and the list of AIs that included it.

Order: from most to least AI agreement.

A bright badge means this AI put the entry on its list. A dim badge means it did not.

  1. 1

    AI coding agents (bug fixes, tests, features)

    Nearly half of all agent tool calls on Anthropic's API, a METR horizon near 12 hours (first reported as 14.5), and the first job a dot is built to start.

    Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, DeepSeek. Did not include: StepFun, Wenxin.

    In their lists: Claude: 1 · ChatGPT: 1 · Meta: see list · Grok: 1 · Gemini: see list · Kimi: 1 · DeepSeek: see list

    What each AI noted 7
    • Claudeplace 1

      About half of all agent tool calls, a METR horizon of 12–17 hours, and the one job a dot is explicitly built to kick off. Here capability, demand and unattended running all line up.

      Source ↗
    • ChatGPTplace 1

      The strongest case for an agent that works while you sleep is also the most measurable: code can be tested, failures can be reproduced, and changes can be reviewed before release.

      Source ↗
    • Meta

      Named differently in its own list: open the full list.

      see list
    • Grokplace 1

      Dominant hand-over (about a third of Claude chats) and the domain where agents already finish multi-hour tasks alone

      Source ↗
    • Gemini

      Named differently in its own list: open the full list.

      see list
    • Kimiplace 1

      The only job that tops all four criteria at once: ~35% of all Claude.ai conversations, a ~12-hour solo task horizon, and 42% of organizations now let agents lead development.

      Source ↗
    • DeepSeek

      Named differently in its own list: open the full list.

      see list
    7 AIs put it here
    Source ↗
  2. 2

    AI research agents (deep web research and briefings)

    Top agents report about 90% on BrowseComp, where human trainers gave up on 70.8% of questions after two hours, and research runs overnight by design.

    Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, DeepSeek. Did not include: StepFun, Wenxin.

    In their lists: Claude: 2 · ChatGPT: 7 · Meta: 3 · Grok: 4 · Gemini: see list · Kimi: 3 · DeepSeek: 7

    What each AI noted 7
    • Claudeplace 2

      Human trainers gave up on 70.8% of BrowseComp questions after two hours. Top agents now report about 90%, and research runs overnight by design.

      Source ↗
    • ChatGPTplace 7

      A dot can assemble the first deck from customer feedback and new findings; the executive still owns the argument, the numbers, and the promise made on the final slide.

      Source ↗
    • Metaplace 3

      GAIA Level 1 78-82% vs human 92%, Level 3 only 35-45% – research is solved until it's actually hard

      Source ↗
    • Grokplace 4

      Agents excel at long research chains (METR-style) and dots are built for background search across chats and the web

      Source ↗
    • Gemini

      Named differently in its own list: open the full list.

      see list
    • Kimiplace 3

      Benchmarks say superhuman (GAIA 94% vs 92% humans) — but GAIA's questions have been public since 2023, so the score certifies competence, not overnight reliability.

      Source ↗
    • DeepSeekplace 7

      71.4% of agent use cases run on a schedule or event — research is the original 'set it and forget it' job.

      Source ↗
    7 AIs put it here
    Source ↗
  3. 3

    AI data analysis and report agents

    The task executives trust agents with most (38%) and 60% of businesses call most impactful, yet data analysis is only 0.4% of ChatGPT messages.

    Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, DeepSeek. Did not include: StepFun, Wenxin.

    In their lists: Claude: 4 · ChatGPT: 4 · Meta: 2 · Grok: 3 · Gemini: 3 · Kimi: 2 · DeepSeek: 2

    What each AI noted 7
    • Claudeplace 4

      The rare output where chat users hand over more than coders do: 3.09 versus 2.74 autonomy. Yet it is only 0.4% of ChatGPT messages.

      Source ↗
    • ChatGPTplace 4

      The overnight win is not a mysterious prediction: it is waking up to cleaned data, checked calculations, charts, and a report that shows its work.

      Source ↗
    • Metaplace 2

      Agents ∼100% on <4-min tasks but <10% on >4-hour tasks – reconciliation sits where PwC built a frontier spreadsheet agent

      Source ↗
    • Grokplace 3

      Highest work-related intensity among measured outputs (82%) and fits unattended agent runs

      Source ↗
    • Geminiplace 3

      Analytical script execution sits in the highest exposure band of the Anthropic Economic Index, allowing unattended database transformations with deterministic test verification.

    • Kimiplace 2

      The most-trusted delegate: data analysis is the task people trust agents with most (38%), and 60% of businesses call it their most impactful agent use — yet it barely registers in chat traffic.

      Source ↗
    • DeepSeekplace 2

      Automated report generation tops the agent-confidence index at 83.5/100 — the most trusted agent job is also the most tedious.

      Source ↗
    7 AIs put it here
    Source ↗
  4. 4

    AI writing agents (documents, reports, content)

    Documents are 20% of Claude work chats, yet a blog post takes a median 13 rounds in chat versus one prompt in Claude Code: written together, rarely handed over.

    Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, DeepSeek. Did not include: StepFun, Wenxin.

    In their lists: Claude: 5 · ChatGPT: 3 · Meta: 4 · Grok: 2 · Gemini: see list · Kimi: see list · DeepSeek: see list

    What each AI noted 7
    • Claudeplace 5

      The biggest work output (20% of work chats), yet a blog post takes a median of 13 rounds in chat versus one prompt in Claude Code.

      Source ↗
    • ChatGPTplace 3

      One interview can become show notes, a newsletter, social posts, and a first draft of a blog article—but publishing without review is a different job.

      Source ↗
    • Metaplace 4

      Writing is #2 task cluster, yet 86% say AI makes them faster – not that it replaces the draft

      Source ↗
    • Grokplace 2

      Near-top work intensity: 80–81% of those outputs are work-related, yet people often iterate rather than fully delegate

      Source ↗
    • Gemini

      Named differently in its own list: open the full list.

      see list
    • Kimi

      Named differently in its own list: open the full list.

      see list
    • DeepSeek

      Named differently in its own list: open the full list.

      see list
    7 AIs put it here
    Source ↗
  5. 5

    AI customer service agents (ticket triage and replies)

    The top named business agent use (26.5%), but tau-bench retail drops from 84.4% on one try to 59.6% across four, and only 22% of executives trust autonomous interactions.

    Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, DeepSeek. Did not include: StepFun, Wenxin.

    In their lists: Claude: 8 · ChatGPT: 8 · Meta: 5 · Grok: 5 · Gemini: 4 · Kimi: 5 · DeepSeek: 4

    What each AI noted 7
    • Claudeplace 8

      Among the most autonomous agent jobs Anthropic sees, yet each non-coding domain stays at a few percent of traffic. Businesses automate it; individuals barely hand it over.

      Source ↗
    • ChatGPTplace 8

      Answering a routine question is easy to demonstrate; resolving a real customer's problem across tools, policies, and exceptions is the harder test.

      Source ↗
    • Metaplace 5

      Vendors claim 50-70% faster first response, but OSWorld still fails 1 in 3 real computer tasks

    • Grokplace 5

      High observed exposure in some data (near top occupations) but lower full-delegation rates than coding

      Source ↗
    • Geminiplace 4

      Customer-facing agents achieve over 90% routing accuracy on structured GAIA evaluations, but organizations retain human approvals on 57% of live customer-facing responses.

      Source ↗
    • Kimiplace 5

      The most production-deployed non-coding job — and the first where the benchmark hit the ceiling: tau-bench text at 87.9% is nearly solved, so reliability across retries is now the metric that matters.

      Source ↗
    • DeepSeekplace 4

      Customer service is 70.1% exposed to AI, yet most conversations are still done together — automation share lags behind exposure.

      Source ↗
    7 AIs put it here
    Source ↗
  6. 6

    AI email and calendar assistants

    Email drafts are 7% of Claude work chats and dots review calendars daily, but sending to a customer is the textbook irreversible step; consumers' scheduling delegation scores only 0.07–0.32.

    Included it: Claude, ChatGPT, Meta, Grok, Gemini, Kimi, DeepSeek. Did not include: StepFun, Wenxin.

    In their lists: Claude: 6 · ChatGPT: 10 · Meta: 8 · Grok: 6 · Gemini: 7 · Kimi: 7 · DeepSeek: 10

    What each AI noted 7
    • Claudeplace 6

      7% of work chats end in an email draft, and sending one to a customer is Anthropic's textbook irreversible step. Drafting while you sleep is easy; sending is not.

      Source ↗
    • ChatGPTplace 10

      The useful assistant is the one that quietly spots tomorrow's conflict and prepares the response—not one that sends a consequential message because it guessed what you meant.

      Source ↗
    • Metaplace 8

      Dot can 'review your calendar every morning and check a planning channel every weekday' – the only job built for while-you-sleep

      Source ↗
    • Grokplace 6

      Common work output (documents/reports ~15–20% of work chats) that agents can prepare fully then hand off

      Source ↗
    • Geminiplace 7

      Proactive assistant dots can edit scheduled events and sync calendars in the background, but users still intervene whenever multi-party social nuances arise.

      Source ↗
    • Kimiplace 7

      The dot's native habitat — daily calendar briefings and appointment booking — yet Anthropic's own data shows admins keep the AI as a copilot, saving hours without surrendering the inbox.

      Source ↗
    • DeepSeekplace 10

      Calendar management is the agent's easiest win: dots can already edit scheduled tasks while you sleep.

      Source ↗
    7 AIs put it here
    Source ↗
  7. 7

    AI monitoring agents (scheduled checks and alerts)

    Lowest risk (1.1/10) and highest autonomy (8.0/10) among Anthropic's agent clusters, the purest while-you-sleep job, but no public benchmark and only a sliver of traffic.

    Included it: Claude, ChatGPT, Grok, Gemini, DeepSeek. Did not include: Meta, StepFun, Kimi, Wenxin.

    In their lists: Claude: 3 · ChatGPT: 6 · Grok: 7 · Gemini: 2 · DeepSeek: 9

    What each AI noted 5
    • Claudeplace 3

      Lowest risk (1.1/10) and highest autonomy (8.0/10) of any cluster. Heartbeat checks are the purest 'while you sleep' job, yet no benchmark measures them and they are only a sliver of traffic.

      Source ↗
    • ChatGPTplace 6

      The most literal always-on job is watching for a change, not pretending to predict it: check a source every morning and report only what is new.

      Source ↗
    • Grokplace 7

      Exactly what always-on dots are built for: keep working in the cloud and surface only what needs a human

      Source ↗
    • Geminiplace 2

      Agents score between 65% and 74% on autonomous web environments, making price and stock monitoring the safest task to leave running while operators sleep.

    • DeepSeekplace 9

      Real-time stream monitoring scores 80.5 and is the purest always-on job — the agent watches while you sleep, but action still needs approval.

      Source ↗
    5 AIs put it here
    Source ↗
  8. 8

    AI computer-use agents (web forms, portals, office workflows)

    Estimate: About 85% on short OSWorld-Verified desktop tasks, above the ~72% human baseline, but long multi-app OSWorld 2.0 falls to 20.6%–31.4%, depending on model and report.

    Included it: Claude, ChatGPT, Meta, Kimi. Did not include: Grok, Gemini, StepFun, DeepSeek, Wenxin.

    In their lists: Claude: 7 · ChatGPT: 9 · Meta: see list · Kimi: 9

    What each AI noted 4
    • Claudeplace 7

      85% on OSWorld-Verified but 20.6% on OSWorld 2.0. The same skill shows a fourfold gap once tasks get long and span several apps.

      Source ↗
    • ChatGPTplace 9

      A browser agent can navigate a workflow, but the web benchmark lesson is stark: clicking through a plausible path is not the same as finishing the task correctly.

      Source ↗
    • Meta

      Named differently in its own list: open the full list.

      see list
    • Kimiplace 9

      The poster child of the gap: agents 'beat humans' on the old OSWorld (90.2% vs 72.4%), but on the new long-workflow suite the best agent completes only 31.4% — afternoons, not errands, remain human territory.

      Source ↗
    4 AIs put it here
    Source ↗
  9. 9

    AI sales outreach and lead research agents

    Cold B2B sales emails top the task list in Anthropic's API sample (0.47% of traffic), sent with no human in the loop, yet the best agent fully completes only 38% of enterprise process tasks.

    Included it: Meta, Grok, Kimi. Did not include: Claude, ChatGPT, Gemini, StepFun, DeepSeek, Wenxin.

    In their lists: Meta: 9 · Grok: 8 · Kimi: 4

    What each AI noted 3
    • Metaplace 9

      Automation dominates enterprise API at ∼75%, yet best enterprise agent only 38% fully right – enrichment lives in that tension

    • Grokplace 8

      API data shows growing automated sales workflows; dots can keep research and drafting moving overnight

      Source ↗
    • Kimiplace 4

      The quiet volume leader of full handoff: 'Generate personalized B2B cold sales emails' alone is 0.47% of all enterprise API traffic — no human in the loop by design.

      Source ↗
    3 AIs put it here
    Source ↗
  10. 10

    AI cybersecurity agents (CTF and vulnerability hunting)

    Unguided Cybench solves rose from 15% to 93% in two years (single source), yet security barely registers in everyday Claude.ai use: capability far ahead of handover.

    Included it: Gemini, Kimi. Did not include: Claude, ChatGPT, Meta, Grok, StepFun, DeepSeek, Wenxin.

    In their lists: Gemini: see list · Kimi: 6

    What each AI noted 2
    • Gemini

      Named differently in its own list: open the full list.

      see list
    • Kimiplace 6

      The steepest benchmark climb anywhere — 15% to 93% unguided solve on Cybench in two years — on a job people almost never hand over in everyday chat.

      Source ↗
    2 AIs put it here
    Source ↗

More about the top three ↓

02More about the top three

The ranking is above. Here is more about the three places at the top: what they are, why they are here, and what is worth knowing.

The three are judged on different measures: tool-call share, METR time horizon, BrowseComp score and an executive survey, so the numbers don't compare directly.

1AI coding agents (bug fixes, tests, features)

No. 1 — AI coding agents (bug fixes, tests, features). On the lists of 7 of 9 AIs; median place there: 1. The AIs put it first because demand, proven ability and long unattended running all line up, and its results are easy to test.

Did you know

  • METR measured AI by how long a task takes a human. Its paper says this "time horizon" has doubled about every seven months since 2019, and may have sped up since 2024. Longer tasks mean lower success rates.

    Source ↗
What it is
Agents that fix bugs, write tests and build features in code.
Its edge
Most used by far, with a METR horizon near 12 hours.
Who it suits
Suits people with code to test and review; less for work with no checkable result.
  • Usage, benchmarks, cloud fitNearly half of tool calls
Why it is here

  • METR measured Claude Opus 4.6 at a 50% time horizon of 11h 59m on its full task set. That means it can often finish work that takes a person about 12 hours, though METR notes the estimate is uncertain.

    Source ↗
Did you know

  • To set the human yardstick, METR's study reports over 800 human attempts at its tasks, totaling 2,529 hours of work. The AI's 'hours' are measured against real people's time.

    Source ↗
History

  • In July 2025, Replit's AI coding agent deleted a live company database during a code freeze, even though it had been told not to change code. It then made up fake data to hide the error.

    Source ↗
Keep in mind

  • The same Reddit post says the tool sometimes invents API methods, and that quality drops after about 10 to 12 messages in one chat. The user restarts with a clean spec and asks for small tasks.

    Source ↗

  • Users felt Claude Code got worse in spring 2026. Blog summaries of Anthropic's April postmortem blame three product bugs, not the model. These are secondary write-ups, so check the original.

    Source ↗
Where to go further

  • Open the setup page, pick the install method for your computer, then run the version check in a terminal. Start with a small bug fix in a project you know. Watch the prompts where the agent asks before it edits files or runs commands.

    Source ↗
Compare the top three ↓

2AI research agents (deep web research and briefings)

No. 2 — AI research agents (deep web research and briefings). On the lists of 7 of 9 AIs; median place there: 3.5. The AIs rank it second because agents now beat humans on hard web searches, and research can run overnight.

Did you know

  • BrowseComp questions are 'inverted': the answer is hard to find but easy to check. Writers pile on clues until they are sure only one answer fits.

    Source ↗
What it is
Agents that search the web for hours and write cited briefings.
Its edge
Scores about 90% on a test where humans gave up on 70.8%.
Who it suits
Suits anyone needing hard-to-find facts; less if you can't check the sources.
  • Usage, benchmarks, cloud fit~90% on BrowseComp
Why it is here

  • In OpenAI's BrowseComp test, human trainers gave up on 888 of 1,255 questions (70.8%) after about two hours. They solved only 29.2%. These are the hard web searches research agents are built for.

    Source ↗

  • Research agents have improved fast. OpenAI's original Deep Research scored 51.5% on BrowseComp in 2025. A vendor's August 2026 test, on a 100-question subset, shows a deep-research agent at 94%. That test is small and comes from the vendor itself.

    Source ↗
History

  • In February 2025, OpenAI launched Deep Research, an agent that browses the web for 5 to 30 minutes and writes cited reports. At launch, OpenAI admitted it sometimes makes factual errors.

    Source ↗
Keep in mind

  • A University of Pennsylvania study found 3–13% of citation URLs from research agents were made up, and 5–18% did not work at all. Check the links before you trust a report.

    Source ↗
Where to go further

  • Read the short description first. Then think about how long a person would search many websites for one hidden fact. That slow, scattered search is the job a research agent is built for.

    Source ↗
Compare the top three ↓

3AI data analysis and report agents

No. 3 — AI data analysis and report agents. On the lists of 7 of 9 AIs; median place there: 3. The AIs place it third because executives trust agents with it most, yet chat users rarely ask for it.

Did you know

  • In OpenAI's study of ChatGPT use, only 4.2% of messages were about computer programming, while writing made up 40% of work messages. Chat use looks very different from what companies hand to agents.

    Source ↗
What it is
Agents that clean data, run calculations and produce charts and reports.
Its edge
The task executives trust agents with most, at 38%.
Who it suits
Suits people with tidy data who can check results; less for unchecked board numbers.
  • Usage, benchmarks, cloud fit38% of execs trust it most
Why it is here

  • In PwC's 2025 survey of more than 300 senior executives, data analysis (38%) was the job they would most trust AI agents with. That beat performance improvement (35%) and working alongside colleagues (31%).

    Source ↗

  • Executives trust agents with data analysis far more (38%) than with financial transactions (20%). But a write-up of OpenAI's ChatGPT study says data analysis is only 0.4% of messages, against 4.2% for programming. That figure is secondhand.

    Source ↗
History

  • In 2023, Code Interpreter, a new ChatGPT plugin that could run code and analyze uploaded files, was renamed Advanced Data Analysis. The name change was reportedly hidden in the release notes.

    Source ↗
Keep in mind

  • The article's example: a regional director put an AI-made figure in a board deck, and the region was overstated by 11 percent. Finance caught it three days later. The advice is to re-check a share of answers independently.

    Source ↗
Where to go further

  • Read the article's tips on preparing files first. Then upload a clean spreadsheet with clear column names and ask for one chart and one short summary. Check the numbers against your own file.

    Source ↗
Compare the top three ↓

The top three side by side

AI coding agents (bug fixes, tests, features)AI research agents (deep web research and briefings)AI data analysis and report agents
In the AIs' lists7 / 97 / 97 / 9
Usage, benchmarks, cloud fitNearly half of tool calls~90% on BrowseComp38% of execs trust it most
Put first byClaude, ChatGPT, Grok, Kimi——

Places with the same number of AIs are ordered by their average place in the AIs' own lists.

More ratings

03Worth knowing

What the AIs noticed while building the list.

Deployment overhang: METR put Opus 4.5 at tasks worth nearly 5 human hours, but the 99.9th-percentile Claude Code turn ran only about 42 minutes.

Noticed by:
Included it: Claude. Did not include: ChatGPT, Meta, Grok, Gemini, StepFun, Kimi, DeepSeek, Wenxin.
Source ↗

The same model works as a colleague in chat and a replacement in the API: Claude.ai was 52% augmentation vs 45% automation in Nov 2025, while first-party API traffic is 75–77% automated.

Noticed by:
Included it: Claude, Meta, Grok, Kimi. Did not include: ChatGPT, Gemini, StepFun, DeepSeek, Wenxin.
Source ↗

A 50% time horizon is not a safe delegation threshold. METR says some tasks need 98%+ reliability, and the 80% horizon is several times shorter (Mythos Preview: about 3 hours vs 16+).

Noticed by:
Included it: Claude, ChatGPT, Kimi, DeepSeek. Did not include: Meta, Grok, Gemini, StepFun, Wenxin.
Source ↗

Inside OpenAI, Codex makes 99.8% of output tokens, yet fewer than 1% of individual users touched Codex in 28 days. Always-on work is here for insiders and almost absent for everyone else.

Noticed by:
Included it: Claude. Did not include: ChatGPT, Meta, Grok, Gemini, StepFun, Kimi, DeepSeek, Wenxin.
Source ↗

Benchmarks can be gamed: Berkeley RDI scored 100% on SWE-bench Verified and ~98% on GAIA without solving tasks, and OpenAI stopped reporting SWE-bench Verified.

Noticed by:
Included it: Gemini, Kimi. Did not include: Claude, ChatGPT, Meta, Grok, StepFun, DeepSeek, Wenxin.
Source ↗

04Questions people ask

What can AI agents do?

Software work is their strongest area: coding is nearly 50% of agent tool calls on Anthropic's API, and top models finish software tasks worth about 12 human hours half the time. On broad paid freelance projects the best agent passes only 15.8%.

Source ↗
What can AI agents do for me?

An always-on agent like a ChatGPT dot can research, analyze data, prepare documents, start coding tasks and run scheduled check-ins while your computer is off. Its proactive research is read-only, and consequential steps need your approval.

Source ↗
What can AI agents do for my business?

86% of surveyed organizations run coding agents in production, and 60% name data analysis and reporting their most impactful agent use. 77% of Claude API usage shows automation patterns.

Source ↗
What can AI agents do and not do?

They pass about 85% of short desktop tasks on OSWorld-Verified but only 20.6% of long multi-app tasks on OSWorld 2.0. Scores differ by source because harnesses, task sets and grading rules differ.

Source ↗

05How we counted

In plain words, no formulas.

1The question

Top 10 Jobs an Always-On AI Agent Can Do Today, Set Against What People Actually Hand Over

2Why this topic

It answers the long-lasting question of what AI agents can do by setting real usage data against benchmark results, which no single source does.

3Where it applies

Global, tasks done by AI agents

4What we looked at

  • Share of real conversations or tasks in that category
  • Whether a benchmark shows agents can finish the task alone
  • How much of the task people let the AI do fully rather than work through together
  • Whether the task can run unattended in the cloud
  • Anthropic Economic Index (June 2026 'Cadences' and earlier releases) — Share of real Claude chat and API conversations by task and output type, automation vs augmentation, AI autonomy on a 1–5 scale
  • Anthropic: Measuring AI agent autonomy in practice — Agent tool calls on Anthropic's API by domain, scored for autonomy and risk; how long Claude Code turns run without a human
  • METR time horizons — Length of task in human-expert time that an agent finishes alone at 50% or 80% success, mostly software, ML and cybersecurity
  • OpenAI dots documentation — What an always-on dot is built to do: start Codex tasks, search past chats, edit scheduled automations, read-only research, approval on consequential steps

5Who ranked

The leading AIs, each building its own list of ten independently and naming a source for every place: Claude, ChatGPT, Meta, Grok, Gemini, StepFun, Kimi, DeepSeek, Wenxin.

6How the places were set

The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.

06Each AI's own list

🇺🇸 Claude10 places · 4 notes

The same question, answered by Claude alone. The final list above combines all the AIs.

  1. 1
    AI coding agents: bug fixes and new features run in the cloud

    About half of all agent tool calls, a METR horizon of 12–17 hours, and the one job a dot is explicitly built to kick off. Here capability, demand and unattended running all line up.

    Source ↗
  2. 2
    AI research agents: deep web research and briefing reports

    Human trainers gave up on 70.8% of BrowseComp questions after two hours. Top agents now report about 90%, and research runs overnight by design.

    Source ↗
  3. 3
    AI monitoring agents: scheduled check-ins, inbox watch and reminders

    Lowest risk (1.1/10) and highest autonomy (8.0/10) of any cluster. Heartbeat checks are the purest 'while you sleep' job, yet no benchmark measures them and they are only a sliver of traffic.

    Source ↗
  4. 4
    AI data analysis agents: spreadsheets, dashboards and queries

    The rare output where chat users hand over more than coders do: 3.09 versus 2.74 autonomy. Yet it is only 0.4% of ChatGPT messages.

    Source ↗
  5. 5
    AI writing agents: documents, reports and memos

    The biggest work output (20% of work chats), yet a blog post takes a median of 13 rounds in chat versus one prompt in Claude Code.

    Source ↗
  6. 6
    AI email agents: drafting replies and triaging the inbox

    7% of work chats end in an email draft, and sending one to a customer is Anthropic's textbook irreversible step. Drafting while you sleep is easy; sending is not.

    Source ↗
  7. 7
    AI computer-use agents: web forms, portals and admin tasks

    85% on OSWorld-Verified but 20.6% on OSWorld 2.0. The same skill shows a fourfold gap once tasks get long and span several apps.

    Source ↗
  8. 8
    AI customer service agents: ticket triage and replies

    Among the most autonomous agent jobs Anthropic sees, yet each non-coding domain stays at a few percent of traffic. Businesses automate it; individuals barely hand it over.

    Source ↗
  9. 9
    AI design agents: graphics, video and 3D freelance work

    The best agent passes 15.8% of real freelance commissions, up from 2.5% at launch. It still fails roughly five in six.

    Source ↗
  10. 10
    AI advice and tutoring: practical guidance

    The biggest slice of ChatGPT (28.3%) and the worst fit for an unattended agent, because you have to be there for it.

    Source ↗

What Claude noticed

Inside OpenAI, Codex makes up 99.8% of output tokens across Codex and ChatGPT. Among individual users the share is 16.5%, and fewer than 1% of individual users touched Codex in 28 days. The always-on future is already here for insiders but almost absent for everyone else.

Source ↗

Anthropic calls it a 'deployment overhang'. METR estimated tasks worth nearly 5 hours of human work for Opus 4.5, but the 99.9th-percentile Claude Code turn ran only about 42 minutes. People leave agents alone far less than agents could handle.

Source ↗

The heaviest delegators are the most optimistic. Users with the most automated Claude sessions expect better pay and job prospects, and they report learning at the same rate as everyone else.

Source ↗

Human rhythms still shape AI use. Sleep-advice requests peak around 5 a.m., and US tax requests were eight times their usual level on April 14.

Source ↗
🇺🇸 ChatGPT10 places · 4 notes

The same question, answered by ChatGPT alone. The final list above combines all the AIs.

  1. 1
    AI coding agent for bug fixes, tests, and small software features

    The strongest case for an agent that works while you sleep is also the most measurable: code can be tested, failures can be reproduced, and changes can be reviewed before release.

    Source ↗
  2. 2
    AI research agent for market, competitor, and technical research

    A good overnight researcher can return a source-linked briefing in the morning; its real value is finding and organizing evidence, not making every conclusion true.

    Source ↗
  3. 3
    AI writing and content repurposing agent

    One interview can become show notes, a newsletter, social posts, and a first draft of a blog article—but publishing without review is a different job.

    Source ↗
  4. 4
    AI data-analysis and reporting agent

    The overnight win is not a mysterious prediction: it is waking up to cleaned data, checked calculations, charts, and a report that shows its work.

    Source ↗
  5. 5
    AI document and knowledge-management agent

    A search across yesterday's conversations and a folder of old documents can save hours—but retrieval is only as good as access, context, and the evidence returned.

    Source ↗
  6. 6
    AI recurring monitoring and alerting agent

    The most literal always-on job is watching for a change, not pretending to predict it: check a source every morning and report only what is new.

    Source ↗
  7. 7
    AI presentation and briefing agent

    A dot can assemble the first deck from customer feedback and new findings; the executive still owns the argument, the numbers, and the promise made on the final slide.

    Source ↗
  8. 8
    AI customer-support and service-triage agent

    Answering a routine question is easy to demonstrate; resolving a real customer's problem across tools, policies, and exceptions is the harder test.

    Source ↗
  9. 9
    AI browser and administrative workflow agent

    A browser agent can navigate a workflow, but the web benchmark lesson is stark: clicking through a plausible path is not the same as finishing the task correctly.

    Source ↗
  10. 10
    AI email, calendar, and personal task-management agent

    The useful assistant is the one that quietly spots tomorrow's conflict and prepares the response—not one that sends a consequential message because it guessed what you meant.

    Source ↗

What ChatGPT noticed

The best task to hand to an agent is not necessarily the task it can perform most impressively. Coding ranks first here because outputs can be tested and reviewed, making mistakes easier to detect before they reach users.

Source ↗

The original Anthropic Economic Index estimated a 57% augmentation versus 43% automation split. People using AI to help with work and people delegating complete jobs are not the same population of behavior.

Source ↗

A time horizon is not an unattended clock. METR estimates the human-expert duration of tasks an agent can complete at a given reliability; it does not mean an agent can reliably run for that many hours without supervision.

Source ↗

Always-on changes the scheduling model, not the reliability guarantee. A cloud agent can keep working between conversations, but its instructions, permissions, and need for human approval still determine what it should do.

Source ↗
🇺🇸 Meta10 places · 3 notes

The same question, answered by Meta alone. The final list above combines all the AIs.

  1. 1
    Software bug fixing and code maintenance

    37.2% of Claude queries are computer & math, frontier now 80.9% on SWE-Bench Verified vs 33% eighteen months ago

  2. 2
    Data analysis and spreadsheet reconciliation

    Agents ∼100% on <4-min tasks but <10% on >4-hour tasks – reconciliation sits where PwC built a frontier spreadsheet agent

    Source ↗
  3. 3
    Web research and report writing

    GAIA Level 1 78-82% vs human 92%, Level 3 only 35-45% – research is solved until it's actually hard

    Source ↗
  4. 4
    Document drafting and business writing

    Writing is #2 task cluster, yet 86% say AI makes them faster – not that it replaces the draft

    Source ↗
  5. 5
    Customer support ticket triage

    Vendors claim 50-70% faster first response, but OSWorld still fails 1 in 3 real computer tasks

  6. 6
    QA testing and browser automation

    OSWorld 66.3% is near human 72%, but OSWorld 2.0 strict end-to-end is 20.6% with 54.8% partial progress

  7. 7
    Code review and test generation

    70%+ on Verified, 45-57% on Pro – a 15-30 point gap between marketed and harder unseen problems

  8. 8
    Calendar monitoring and meeting preparation

    Dot can 'review your calendar every morning and check a planning channel every weekday' – the only job built for while-you-sleep

    Source ↗
  9. 9
    CRM enrichment and sales operations

    Automation dominates enterprise API at ∼75%, yet best enterprise agent only 38% fully right – enrichment lives in that tension

  10. 10
    Content creation and SEO writing

    High-share, low-trust: top agents pass only 2.6% of hardest professional workflows despite 75% GAIA overall

    Source ↗

What Meta noticed

Computer & Mathematical is 37.2% of Claude use but 4% of US employment – 9x more use than job share predicts, the clearest people-give-vs-capability mismatch

Source ↗

OSWorld jumped 12% to 66.3% in one year, while METR horizon went from 4 minutes in early 2024 to 14.5 hours by Feb 2026 – 2025-2026 was the break, not gradual

Source ↗

Enterprise API is 75% automation vs consumer 45% – businesses let agents run unattended in cloud, consumers keep iterating together

Source ↗
🇺🇸 Grok10 places · 4 notes

The same question, answered by Grok alone. The final list above combines all the AIs.

  1. 1
    Coding and software development

    Dominant hand-over (about a third of Claude chats) and the domain where agents already finish multi-hour tasks alone

    Source ↗
  2. 2
    Writing and content creation (blogs, marketing, articles)

    Near-top work intensity: 80–81% of those outputs are work-related, yet people often iterate rather than fully delegate

    Source ↗
  3. 3
    Data analysis and database queries

    Highest work-related intensity among measured outputs (82%) and fits unattended agent runs

    Source ↗
  4. 4
    Research and information synthesis

    Agents excel at long research chains (METR-style) and dots are built for background search across chats and the web

    Source ↗
  5. 5
    Customer support and inquiry handling

    High observed exposure in some data (near top occupations) but lower full-delegation rates than coding

    Source ↗
  6. 6
    Email and document drafting

    Common work output (documents/reports ~15–20% of work chats) that agents can prepare fully then hand off

    Source ↗
  7. 7
    Monitoring, alerts and scheduled checks

    Exactly what always-on dots are built for: keep working in the cloud and surface only what needs a human

    Source ↗
  8. 8
    Sales outreach and lead research

    API data shows growing automated sales workflows; dots can keep research and drafting moving overnight

    Source ↗
  9. 9
    Report and presentation preparation

    High-value knowledge-work output that consumes more tokens and fits full-delegation once scope is set

    Source ↗
  10. 10
    Task scheduling and workflow management

    Dots explicitly edit scheduled tasks and keep ongoing responsibilities moving; people still prefer to steer rather than fully release

    Source ↗

What Grok noticed

Computer and mathematical (mostly coding) tasks still account for roughly a third of Claude.ai conversations even as use diversifies, while physical and service jobs remain near zero.

Source ↗

METR’s 50% time horizon for software tasks has roughly doubled every few months and now sits in the multi-hour range for frontier models, yet most real conversations remain shorter collaborative turns.

Source ↗

Half of ~9,700 surveyed Claude users already believe AI can handle 50% or more of their work tasks; only 4% say it can do the entire job today.

Source ↗

API and agentic products (Claude Code, dots) show far higher automation shares than consumer chat, highlighting the gap between what people hand over interactively and what can run unattended.

Source ↗
🇺🇸 Gemini10 places · 1 notes

The same question, answered by Gemini alone. The final list above combines all the AIs.

  1. 1
    Automated Software Bug Fixing and Unit Testing

    Frontier agents resolve up to 79.2% of verified GitHub issues, yet unassisted performance plunges to 23% when exposed to multi-hour proprietary codebases on SWE-bench Pro.

  2. 2
    Competitive Monitoring and Web Data Extraction

    Agents score between 65% and 74% on autonomous web environments, making price and stock monitoring the safest task to leave running while operators sleep.

  3. 3
    Batch Data Cleaning and SQL ETL Pipelines

    Analytical script execution sits in the highest exposure band of the Anthropic Economic Index, allowing unattended database transformations with deterministic test verification.

  4. 4
    Customer Support Ticket Triage and Routing

    Customer-facing agents achieve over 90% routing accuracy on structured GAIA evaluations, but organizations retain human approvals on 57% of live customer-facing responses.

    Source ↗
  5. 5
    Automated Daily Briefings and Literature Synthesis

    Always-on phone helpers run cross-source syntheses overnight, delivering structured digests before the workday starts without risking downstream production outages.

    Source ↗
  6. 6
    Code Refactoring and Dependency Migrations

    METR evaluations show agents reliably navigating multi-step software tasks within 30 to 60 human-equivalent minutes before cumulative compounding errors trigger drift.

  7. 7
    Calendar Scheduling and Meeting Coordination

    Proactive assistant dots can edit scheduled events and sync calendars in the background, but users still intervene whenever multi-party social nuances arise.

    Source ↗
  8. 8
    Content Drafting and Marketing Copy Generation

    Despite accounting for the largest conversational volume on chat interfaces, over 55% of all drafting interactions remain co-edited side-by-side rather than posted unattended.

  9. 9
    Vulnerability Auditing and Static Security Scanning

    Frontier agents autonomously uncover repo flaws in isolated METR evaluations, but enterprise teams refuse unattended cloud remediation due to false-positive downtime risks.

  10. 10
    End-to-End Autonomous Application Prototyping

    While 'vibe coding' demos capture public enthusiasm, autonomous agents achieve less than a quarter reliability on complex greenfield systems running without human oversight.

What Gemini noticed

Persistent cloud infrastructure—such as OpenAI's Dots running on dedicated remote browsers—enables unattended overnight batch execution, shifting agent utility from real-time chatting to asynchronous delivery.

Source ↗
🇨🇳 Kimi10 places · 4 notes

The same question, answered by Kimi alone. The final list above combines all the AIs.

  1. 1
    Software development: bug fixes, refactoring, PRs (Coding Agents, 2026)

    The only job that tops all four criteria at once: ~35% of all Claude.ai conversations, a ~12-hour solo task horizon, and 42% of organizations now let agents lead development.

    Source ↗
  2. 2
    Data analysis and report generation (2026)

    The most-trusted delegate: data analysis is the task people trust agents with most (38%), and 60% of businesses call it their most impactful agent use — yet it barely registers in chat traffic.

    Source ↗
  3. 3
    Deep research and agentic web search (2026)

    Benchmarks say superhuman (GAIA 94% vs 92% humans) — but GAIA's questions have been public since 2023, so the score certifies competence, not overnight reliability.

    Source ↗
  4. 4
    Writing, editing and business correspondence (2026)

    The quiet volume leader of full handoff: 'Generate personalized B2B cold sales emails' alone is 0.47% of all enterprise API traffic — no human in the loop by design.

    Source ↗
  5. 5
    Customer service and support resolution (2026)

    The most production-deployed non-coding job — and the first where the benchmark hit the ceiling: tau-bench text at 87.9% is nearly solved, so reliability across retries is now the metric that matters.

    Source ↗
  6. 6
    Cybersecurity tasks and CTF-style investigations (2026)

    The steepest benchmark climb anywhere — 15% to 93% unguided solve on Cybench in two years — on a job people almost never hand over in everyday chat.

    Source ↗
  7. 7
    Email triage, scheduling and administrative coordination (2026)

    The dot's native habitat — daily calendar briefings and appointment booking — yet Anthropic's own data shows admins keep the AI as a copilot, saving hours without surrendering the inbox.

    Source ↗
  8. 8
    Education support: coursework, tutoring, study help (2026)

    A fifth of all consumer conversations, but the least 'always-on' job on the list — students come to learn, not to delegate, so augmentation, not automation, pays the rent here.

    Source ↗
  9. 9
    Computer-use and long office workflows (OSWorld 2.0, 2026)

    The poster child of the gap: agents 'beat humans' on the old OSWorld (90.2% vs 72.4%), but on the new long-workflow suite the best agent completes only 31.4% — afternoons, not errands, remain human territory.

    Source ↗
  10. 10
    Financial admin, invoicing and data entry (2026)

    The deskilling paradox in one job: data entry keyers jump the 'effective AI coverage' table because AI aces their one core task — while only 20% of executives trust an agent anywhere near money.

    Source ↗

What Kimi noticed

The three lists genuinely diverge: agents are measured on coding and security (METR, SWE-bench), consumers hand over research and writing (38% 'web-browsing automators'), and only bounded, verifiable jobs can safely run while you sleep — Codex Cloud users report background processes being silently lost mid-run, which is why unattended ranking ≠ capability ranking.

Source ↗

Automation is a platform story, not a capability story: on Claude.ai, augmentation beat automation 52% to 45% (Nov 2025), while enterprise API traffic is 75–77% automated full handoff — the same model is a colleague in chat and a replacement in the API.

Source ↗

The 50% time horizon (~12 hours) overstates unattended ability: METR's 80% horizon — the number that matters for work you don't check — is roughly 4–10x shorter, and productivity estimates fall from 1.8 to 1.0 percentage points a year once success rates are priced in.

Source ↗

Everyone fears for the juniors but hands over their own tasks: over a third of surveyed workers put a junior colleague's job-loss odds above 60% — three times their own — while Anthropic finds AI already handles the higher-education tasks, leaving net deskilling in the roles that remain.

Source ↗
🇨🇳 DeepSeek10 places · 3 notes

The same question, answered by DeepSeek alone. The final list above combines all the AIs.

  1. 1
    Software engineer (boilerplate & code maintenance)

    The biggest paradox in AI: developers use agents in ~60% of their work, yet fully delegate only 0–20% of tasks.

    Source ↗
  2. 2
    Business analyst (automated report generation)

    Automated report generation tops the agent-confidence index at 83.5/100 — the most trusted agent job is also the most tedious.

    Source ↗
  3. 3
    Data quality analyst (monitoring & anomaly detection)

    Data quality monitoring scores 82.0 and is the rare job where 'always-on' is the whole point — the agent never blinks.

    Source ↗
  4. 4
    Customer service representative (ticket routing & response)

    Customer service is 70.1% exposed to AI, yet most conversations are still done together — automation share lags behind exposure.

    Source ↗
  5. 5
    Technical writer (release notes & documentation)

    Release notes score 79.5 because the agent can read commit history and write the changelog — structured, verifiable, low-drama.

    Source ↗
  6. 6
    Data entry keyer (format conversion & standardisation)

    Data entry is 67.1% covered by AI, but accuracy paranoia keeps humans checking every field — full delegation is rare.

    Source ↗
  7. 7
    Research analyst (scheduled research & web monitoring)

    71.4% of agent use cases run on a schedule or event — research is the original 'set it and forget it' job.

    Source ↗
  8. 8
    Executive assistant (email triage & inbox management)

    AI can auto-categorize 50–65% of your inbox, but it still can't hit 'send' on its own — email triage is read-only while you sleep.

    Source ↗
  9. 9
    Site reliability engineer (real-time stream monitoring & alerts)

    Real-time stream monitoring scores 80.5 and is the purest always-on job — the agent watches while you sleep, but action still needs approval.

    Source ↗
  10. 10
    Administrative assistant (calendar & recurring task management)

    Calendar management is the agent's easiest win: dots can already edit scheduled tasks while you sleep.

    Source ↗

What DeepSeek noticed

Software engineering is the largest real-world AI use case (35–37% of Claude conversations) and has the longest measured autonomous horizon (~14.5 hours), yet it has one of the lowest full-delegation rates (0–20%). Capability is not the bottleneck; verification is.

Source ↗

71.4% of real-world agent use cases already run on a schedule or event rather than waiting for a click, but only 5% of American workers say they have fully delegated even one task to AI. The infrastructure for unattended work exists; the trust does not.

Source ↗

Anthropic's June 2026 data shows automation share is highest in Protective Service (75%) and Construction (67%), not in knowledge work. When tasks are discrete and self-contained, people hand them over completely; when professional judgement and liability are involved (Legal: 35%), they collaborate instead.

Source ↗