The agent time horizon is doubling

You hand an agent a ticket and go get coffee. Ten minutes is about the right length of leash: long enough for it to do something useful, short enough that you can read the whole transcript when you sit back down. That leash is the reason most teams still treat an agent as a faster autocomplete. You watch it because watching is cheap.

The leash is getting longer, and the measurement is public. When it reaches a day, reading the whole transcript stops being a thing a person does. When it reaches a week, the practices built around watching are already gone, whether or not anyone replaced them.

#What METR measures

Kwa and colleagues proposed a metric that translates a benchmark score into something an operator can reason about: the 50 percent task-completion time horizon, defined as the time a human with relevant domain expertise takes on tasks the model finishes with a 50 percent success rate.¹ They timed people on RE-Bench, HCAST, and 66 shorter tasks they wrote for the purpose, then scored models against the same set.

Two findings came out of it. Frontier models at the time of writing sat at a 50 percent horizon of roughly 50 minutes. And the frontier horizon had been doubling approximately every seven months since 2019.

METR revised the dataset in January 2026.² The task suite grew to 228 tasks, and the number of tasks taking eight hours or more went from 14 to 31. The revised trend puts the doubling at 196.5 days across the whole period and 130.8 days across models released since 2023. The highest measured 50 percent horizon in that release was 320 minutes, with a confidence interval running from 170 to 729 minutes.

The caveats travel with the number, and METR states them plainly. Human baseline times were measured for only 5 of the 31 long tasks; the rest are estimates. The confidence intervals are wide. The public tracker carries a notice that measurements above 16 hours are unreliable with the current task suite.³ A metric that reports its own ceiling is a better metric than one that does not, and the ceiling is the part worth keeping: nobody has a trustworthy measurement of an agent working for a week, because nobody has a task suite that can grade one.

#The arithmetic, and what it costs to extrapolate it

A doubling time is a small number that behaves badly in a planning conversation. Three doublings sounds like a rounding error and is a factor of eight.

Read that as arithmetic and not as a roadmap. A line that has held for seven years can stop in the eighth, the confidence intervals on the recent points are wide enough to move every row, and the grading problem arrives before the capability does. What the table is good for is deciding how much of your current practice depends on the leash staying short. If the answer is most of it, your planning horizon is shorter than the capability horizon, which is the uncomfortable part.

#The 50 percent line is not the production line

Half is a strange place to draw a line if you are the one shipping the result. It is the right place for a research metric, because it is the steepest part of the curve and the most sensitive to a change in the model. It is the wrong place for a deployment decision, because the thing you are deciding is whether to let a run finish while you sleep.

Yao and colleagues made the same point with a different metric. Standard practice reports pass@k, the probability that at least one of k attempts succeeds. They define pass^k instead: the probability that all k independent attempts at a task succeed.⁴ State-of-the-art function-calling agents in their study succeeded on under 50 percent of tasks, and pass^8 in the retail domain fell below 25 percent. Sampling more attempts raises the first number and lowers the second, so a system tuned on pass@k can get less consistent while its headline improves.

A five-hour horizon at 50 percent success means the agent gets a five-hour task right about as often as a coin lands heads. Run one such task a day for a week and the chance that all five land is around 3 percent. That is not an argument against long runs. It is an argument that the horizon number tells you what an agent can reach, not what you can ship, and those are two different purchases.

#Long runs do not fail the way short runs fail

The obvious theory is that long runs fail because the context window fills up. The measurement says otherwise.

Backlund and Petersson built Vending-Bench to test coherence rather than capability: an agent runs a vending machine business, managing inventory, placing orders, setting prices, and covering daily costs.⁵ Each task is simple. The run is not. Individual runs exceed 20 million tokens. What they found is that agents misread delivery schedules, forget orders they placed, and fall into what the authors call tangential meltdown loops from which they rarely recover. They report no clear correlation between failures and the point at which the context window fills, and high variance between runs of the same model on the same task.

Sinha and colleagues name a mechanism that fits. Studying long-horizon execution, they find a self-conditioning effect: a model becomes more likely to make mistakes when its context already contains its own errors from earlier turns.⁶ The failure is not that the agent forgot. It is that the agent read its own bad work and treated it as settled. Their other finding is the encouraging half of the same paper: small gains in single-step accuracy compound into large gains in the length of task a model can complete, which is why the horizon curve moves at all.

Put those together and the shape of a long run is clear enough. It does not degrade smoothly toward the end. It proceeds correctly, takes one wrong turn, writes that turn into its own record, and then spends hours being consistent with it. By the time a person looks, the wrong turn is forty steps back and everything after it is coherent.

#Supervision does not scale by watching

Most teams supervise an agent the way they supervise a new engineer, because that is the process they already had.

The loop closes because a person is present for the middle two steps. Every property the team relies on comes from that presence: the wrong turn gets caught early, the correction is cheap, and the reviewer has watched enough of the run to know what the diff means. None of those properties survive a horizon the reviewer cannot sit through. The loop does not break loudly. It keeps running with the watching step quietly skipped, and the review step now reads a week of work with no idea which forty steps mattered.

What replaces presence is not more attention. It is deciding in advance what you used to decide in the moment: what the agent may reach, what it may spend, what it is told before it starts, and which requests stop and wait for a person. Those are the same decisions, moved from the transcript to the mandate. The next posts in this pillar take them one at a time.

#Where this meets Oxagen

Oxagen is workforce management for autonomous agents. It does not run them. What it holds is the mandate: the identity each agent acts as, the systems and data it may request, the budget and rules it runs under, the tools and skills it is equipped with, and the record of what it did.

Three parts of that bear directly on a longer horizon. Requests are answered at the moment of use, so a run that reaches for a system at hour forty gets the same rule it would have got at minute one, and a rule can allow, deny, or route the request to a named person while the run waits. Every governed action is priced and attributed to the person, the agent, the run, the turn, and the step, so a week-long run has a cost you can read per step rather than one line on a monthly bill. And the record keeps each run frame by frame against its mandate, so finding the wrong turn is a query rather than a reread.

That applies to actions routed through Oxagen. A call that does not pass through it is not governed by it, and no record makes an agent correct. What the record does is make a long run something a person can inspect afterward, which is the property the watching step used to provide.