The AI Spotlight
AI Agents

AI Agents Are Doubling Every 7 Months. This One Chart Changes Everything

Researchers at METR found a simple way to measure AI progress, and the answer is a doubling curve. Here is what it measures, why it matters, and the caveats nobody puts in the thumbnail.

A
The AI Spotlight DeskNov 6, 2026 · 4 min read
Advertisement
AdSense slot · articleTopplaceholder until ca-pub ID is set

Ask ten people whether AI is getting better and you will get ten answers shaped by whatever demo they saw last. Benchmarks do not help much either: most of them are quizzes that models either ace or get saturated within a year, which tells you little about whether AI can actually do your job.

In March 2025, the nonprofit research group METR proposed a cleaner yardstick. Instead of asking “what score did the model get,” they asked: how long a task, measured in human working time, can an AI agent complete on its own? The answer they found, published as “Measuring AI Ability to Complete Long Tasks”, has become one of the most shared charts in AI.

What METR actually measured

The team assembled a set of software engineering and research tasks, then timed skilled humans doing them. Some took a few seconds. Some took many hours. That gave every task a “human time” label.

Then they ran AI agents built on frontier models released between 2019 and early 2025 on the same tasks, and recorded which ones each model completed. For each model they computed a 50% time horizon: the length of task, in human minutes, that the model completes successfully about half the time.

Advertisement
AdSense slot · articleMidplaceholder until ca-pub ID is set

The headline result, from the paper on arXiv: over roughly six years, that time horizon has been doubling about every seven months. Early models could only finish tasks that took a person seconds. The strongest model METR tested in early 2025 had a 50% time horizon of roughly an hour.

Why a doubling curve is a big deal

Doubling is the kind of growth human intuition handles badly. If the trend simply continues, the arithmetic goes like this:

  • One hour, then two hours seven months later
  • A full workday a bit over a year after that
  • A full work week around two years further on
  • A month-long project a couple of years beyond that

METR itself flagged that if the trend held, AI agents might handle many software tasks that take humans a month within years rather than decades. That is the part that makes the chart feel scary: it turns “AI is improving” into a rough timeline you can argue about.

It also explains a lot of the last two years of AI news. The shift from chatbots to agents, the rush to ship coding agents, and the explosion of “AI does an entire task for you” demos all sit on this curve. Labs are building for longer tasks because the models can increasingly sustain them.

The caveats that matter

A good chart deserves honest fine print, and METR provided plenty of it.

It is mostly software tasks. The task suite leans heavily on coding and technical research, the domain where models are strongest and results are easiest to check automatically. Whether the same slope holds for sales, design, law or management is an open question.

50% is not reliable. A model that succeeds half the time on hour-long tasks is not a model you can delegate hour-long tasks to. METR also reported an 80% time horizon, and it is several times shorter than the 50% number. For real work, the reliability you need is often higher still.

Benchmarks are cleaner than jobs. Real tasks come with vague instructions, missing context, office politics and messy systems. METR noted that its tasks are more self-contained than typical work, and that models did worse on “messier” tasks.

Trends can bend. Six years of a straight line on a log chart is impressive, but extrapolating any curve is a bet, not a law. Compute costs, data limits or new breakthroughs could slow it down or speed it up.

How to use the 7-month rule in real life

Even with the caveats, the framing is genuinely useful if you build things or plan a career.

  1. Measure tasks in time, not difficulty. When deciding what to hand to an AI, ask how long it would take a competent person. Short tasks are increasingly automatable; long ones still need you to break them up.
  2. Decompose ruthlessly. If agents are good at one-hour chunks, your job becomes turning a two-week project into well-specified one-hour pieces and checking the output.
  3. Expect the ceiling to move. Something an agent cannot do reliably today may be routine in a year or two. Revisit your “AI can’t do that” list every few months.
  4. Value verification skills. As agents take on longer tasks, the bottleneck shifts to whoever can tell good work from plausible-looking garbage.

A companion finding worth knowing

METR is also behind a study that complicates the hype: in a July 2025 randomized trial with experienced open-source developers, those allowed to use AI tools took longer to finish real tasks, even though they believed they were faster. We break that one down in a separate story. The two results are not contradictory. Capability on clean tasks can grow fast while real-world productivity lags, because real work is messy in exactly the ways benchmarks avoid.

The bottom line

The 7-month rule is not a prophecy. It is a measurement, with real limits, of one specific thing: how long a task AI agents can finish on their own about half the time. But it is the clearest evidence yet that agent capability is compounding, not plateauing. If you are planning anything that depends on what AI can and cannot do two years from now, this is the chart to keep on your wall, along with its fine print.

Sources

Up next · 4 min readDevelopers Swore AI Made Them 20% Faster. The Stopwatch Said 19% Slower.A randomized trial with experienced open-source developers found AI tools slowed them down, even as they felt sped up. Here is what the study really shows and how to avoid the trap.Keep reading →

More stories you'll want to read