The Exponent Borrows Two Things Nobody Has Measured

Every agent deployment estimate runs the same arithmetic. Odds a single step goes right, raised to the number of steps. Watch the floor open. Ninety-nine percent per step is a coin flip by step seventy. Unsentimental, directionally right, and I have never seen anyone argued out of it. Nor should they be.

It borrows two things the evidence hasn't supplied.

The first is one per-step number to raise. Towards a Science of AI Agent Reliability, accepted at ICML 2026, takes reliability apart into twelve metrics across four dimensions (consistency, robustness, predictability, safety), evaluated over fourteen agentic models on two benchmarks. What comes back has a shape rather than a value. Two models can sit at identical accuracy and disagree about everything underneath — whether their failures repeat, whether they survive a reworded instruction, whether they have any inkling of which tasks they are about to botch. Capability, the authors note, has been pulling ahead of reliability.

The scorecard itself involves a judgment call. Three dimensions get averaged into a headline figure, and safety is left out on purpose, because folding it in would iron over rare severe failures. Averaging is lossy compression, and somebody has to rule in advance on which losses are affordable. The same average caseload can be a steady trickle or three explosive clusters. You would staff those two worlds nothing alike.

Four of the twelve metrics go straight at spread, every task run five times. Outcome consistency came back low: what an agent solved once, it often missed on the retry. The kinds of actions it took stayed fairly stable across runs. The order did not. Cost, wall-clock time and API calls swung substantially from run to run of the same task.

Same task, same agent, same instruction, different bill.

Which raises the second borrowed assumption: that steps fail independently. A July synthesis across 27 papers and 19 benchmarks proposes a no-recovery bottleneck. Agents seldom catch an early mistake and back out of it, so one early divergence bends everything after it, and the odds of a long run finishing clean fall away faster than flat per-step odds predict. The naive exponent is already a decay curve; the finding says it is also biased, and biased optimistic, in a direction no amount of shaving your per-step estimate will fix. The synthesis is unreviewed, and correlated failure there is a proposed mechanism, not a measured parameter. Its compositional finding is blunt anyway: competence at sub-skills does not reliably add up to end-to-end success, and piling on tools, agents, reasoning effort or context does not uniformly help. It cites METR at near-total success on tasks a skilled human finishes in roughly four minutes, and under ten percent past roughly four hours.

Horizon is doing most of the damage in these results, and horizon is the variable measured least directly. Neither paper isolates task length experimentally. The ICML study lets difficulty stand in for length, which is not the same thing, and what comes back is no tidy widening: accuracy drops, resource consistency degrades, robustness shows no pattern at all, outcome consistency goes mixed. A hard enough task fails with impressive consistency.

So the shape of the consistency curve over a long session remains unmapped. Nothing tests whether one context strategy holds steadier than another across twenty steps. The ICML authors say as much about their own field: evaluations today run for minutes, and multi-hour work needs benchmarks nobody has built.

The number in your deck describes runs that already finished. The run in flight is a draw from a distribution nobody has characterized at that length.

What does a long-running task have to carry with it, as it runs, for anyone to tell how it is going?