Agent task length doubles every seven months. The reliable version is 18 months behind.

The METR Data

METR has been tracking one number since 2019, and it is not a benchmark score. It is a duration: the length of a task, measured in how long a human expert needs, that a frontier agent can finish on its own.

The headline result from their NeurIPS 2025 paper is that this duration has been doubling roughly every seven months since 2019, across a suite of about 230 tasks. Over 2024 and 2025 the doubling compressed to about four months.

That is the curve everyone quotes. It is measured at a 50% success rate. METR measures a second curve at 80%, and that one sits several times lower. Their figures put the 80% horizon around four to six times shorter than the 50% horizon, with both curves climbing at close to the same rate.

Run the arithmetic. Four to six times is between two and two and a half doublings. At seven months per doubling, the reliable curve is trailing the famous one by roughly a year and a half. Every real decision you make about putting agents to work lives in that gap.

The Gap Is Not a Model Problem

It is tempting to read the two curves as "the model is not good enough yet, wait for the next release." That reading has cost people a lot of money.

Look at what actually separates a 50% run from an 80% run on the same task:

• The model did not get dumber on the failures. It hit an input shaped slightly differently than the ones it got right.
• It made a defensible choice at a fork where the correct answer was never written down.
• It finished, produced something plausible, and nobody could tell from the output whether this was one of the good runs.

None of that is fixed by a better model. It is fixed by narrowing the job until the ambiguity is gone, and by deciding in advance what happens on the runs that miss. That work is not prompt writing. It is specification, and it is the entire discipline that agent operations turns out to be.

Here is the thing worth internalizing: the 50% curve tells you what is now possible. The 80% curve tells you what you can staff. Confusing the two is how a demo that impressed everyone in the room becomes an incident three weeks later.

Sort the Job Before You Schedule It

The practical version of this fits in one question. Before an agent job goes on a schedule, ask what a wrong run costs, and put the job in one of three boxes.

Box one: it finishes. A wrong run is cheap, visible, and reversible. Nobody outside sees it, and the next run overwrites it. Examples include:

• Research passes
• Drafts
• Classification into a queue
• Monitoring that reports rather than acts
• Anything that writes to a file you own

These are the jobs to let run unattended, and there are far more of them than most people schedule. A 50% agent is genuinely useful here, because the cost of the misses rounds to zero and the hits compound.

Box two: it drafts, you commit. A wrong run is recoverable but embarrassing or expensive to undo. Anything that leaves your control:

• Outbound messages
• Published copy
• Changes to a paying customer's data
• Spending

The agent does everything up to the irreversible step and then stops. This is where the 80% curve actually bites, because 80% is not good enough to send on its own and it is far more than good enough to do the ninety percent of the work that precedes sending. Separating capture from send is the single highest-leverage split in this whole field, and almost nobody makes it explicitly.

Box three: it does not touch this. A wrong run cannot be undone, or the blast radius is not bounded. Examples include:

• Deletion
• Key rotation
• Anything financial
• Anything that changes access

The right answer is not a more careful prompt. It is that the job is not an agent job this year.

Two rules make the triage hold up:

1. Bound every job in the first two boxes with a window and a budget, so a stuck run costs you a known amount and then stops.
2. Make the deliverable a record rather than prose: a row you can count and audit later, not a paragraph telling you it went well.

An agent that reports its own success in sentences is unfalsifiable, and unfalsifiable output is how you end up trusting a system that quietly stopped working.

The Boxes Move. The Procedure Does Not.

This is why the triage is worth learning rather than waiting out. As the reliable curve climbs, jobs migrate from box three to box two, and from box two to box one. That migration is the whole story of the next few years, and it is genuinely fast: a year and a half of lag is not a decade.

But the sorting question never changes, because it was never a question about the model. It is a question about what a wrong answer costs you, and that is a property of your business. Which means the person who learns to sort now is not doing throwaway work. They are building the thing that stays useful while everything underneath it gets better.

Where the Paying Work Goes

Generation is close to free and getting freer. The scarce thing is somebody who can:

• Look at a real business
• Name the jobs
• Sort them into those three boxes
• Write the specification tight enough that the agent lands in the reliable band
• Keep the thing running when reality changes shape

That role has a name now, and it is showing up in job postings as forward deployed engineer. Strip away the title and it is the same trade: the value has moved from producing software to installing and operating it.

Companies are not short of output. They are short of anyone who will take responsibility for a running system. If you can do that, you are not competing with a model. You are the reason the model is worth anything to a business.

The Invitation

The Agent Ops Club is built around exactly this. The framing on the page is "free to hire, yours to run": you hand an AI employee the work you are furthest behind on, and it runs that job on a schedule. It is the box-one and box-two idea turned into something you can actually put to work rather than read about. There is a free tier, so you can go see whether the argument above survives contact with your own backlog before you spend anything.

If you take one thing from this post and skip the link entirely, take the triage. Write down your three boxes this week, honestly, and notice how many jobs you have been keeping in box three out of habit rather than blast radius.

More of this is written up with the builders in the Vibe Coding is Life community, and I would genuinely like to see other people's box-two lists, because that is where I think most of the unclaimed leverage is sitting right now.

So the small ask: what is one job you would move into box one today if you trusted it to fail quietly? Leave it in the comments and I will tell you how I would bound it.