Why do averages fail for AI agent fleets?

Averages fail for agent fleets because agent cost, latency, and failure blast-radius follow power-law distributions, not Gaussian ones. For a power law with exponent γ ≤ 3 the second moment diverges: the variance is effectively infinite, so the mean is finite but describes almost nothing about any individual run. Production LLM systems show 35x gaps between median and P99 latency, and roughly 3% of tenants consuming 60% of tokens. The practical consequence is that interventions derived from averages — capacity planned on mean load, budgets forecast on mean cost, optimization aimed at the typical run — target a run that does not exist, while the tail carries the consequences.

Averages Lie: Agent Fleets Live in Extremistan

Your agent dashboard reports a mean. The mean is real. It also describes nothing.

Here is the number that should end the argument. Barabási measured the World Wide Web's link structure and found an average in-degree of 4.60. A reasonable person reads that and pictures a typical page with four or five inbound links.

Three quarters of the population sits below the average. The average is not the middle of anything. Barabási writes the expectation for a randomly chosen document as k = 4.60 ± ∞ — a mean with an infinite error bar, which is his way of saying the number is arithmetically correct and analytically useless.

This is not a quirk of the web. It is what happens whenever γ, the power-law exponent, falls at or below 3:

”

"For many scale-free networks the degree exponent γ is between 2 and 3. Hence for these in the N → ∞ limit the first moment ⟨k⟩ is finite, but the second and higher moments, ⟨k²⟩, ⟨k³⟩, go to infinity." — Barabási, Network Science (p. 135)

The second moment is the variance. When it diverges, every tool built on finite variance breaks: standard deviation, confidence intervals, the t-test, capacity planning from mean load, cost forecasting from mean spend. Not "becomes less accurate." Breaks — in the sense of producing predictions of the wrong qualitative type.

Agent fleets are not in Mediocristan

Taleb's name for the two regimes is more useful than the math notation. Mediocristan is where no single observation moves the total. Extremistan is the opposite:

”

"Extremistan: a process where the total can be conceivably impacted by a single observation (say, income for a writer). Also called 'fat-tailed.' Includes the fractal, or power-law, family of distributions." — Taleb, Skin in the Game

Human height is Mediocristan — the tallest person alive does not shift the mean. Agent token spend is Extremistan: one runaway loop can exceed the rest of the month. The 2026 production data is unambiguous.

That last row is the one that matters for anyone running multi-agent swarms. Multiplicative processes are the classic generator of power laws, and a five-step agent chain is multiplicative by construction. Each step's variance multiplies rather than averages out. Adding agents does not smooth the distribution — it fattens the tail.

What the mean actually breaks

Three specific failures follow, and each is a wrong-regime error rather than an imprecision.

Capacity and budget planning. Provisioning against mean load assumes the tail is bounded by a few standard deviations. Under a divergent second moment there is no such bound. The budget is not slightly optimistic; it is derived from a parameter that does not constrain the outcome.

SLOs written on the median. A P50 of 800ms alongside a P99 of 28 seconds means the median describes a run that is not representative of the experience generating your support tickets. Optimizing the median can leave the tail untouched or make it worse.

Outlier removal. This is the most damaging habit, because it is taught as hygiene. Standard practice says identify outliers and exclude them before analysis. Barabási notes the temptation directly — and the cost:

”

"removing these 'outliers' could be seen as data manipulation; if kept, however, one cannot detect the statistical significance of the power law fit." — Barabási, Network Science (p. 173)

In a fat-tailed distribution the outliers are not noise around the signal. They are the signal. Taleb's technical appendix makes the mathematical version of the point: for subexponential distributions, when the sum exceeds some large value, that sum is dominated by the single maximum observation while everything else contributes negligibly. Strip the top runs out of your cost data and you have deleted the majority of your actual spend, then computed statistics on what remains.

Averaging interventions fail structurally

This is where the argument stops being about measurement and starts being about what you do on Monday.

Barabási's robustness result is the cleanest demonstration. Scale-free networks are simultaneously the most robust and the most fragile networks known, depending entirely on the type of perturbation:

”

"while random node failures do not fragment a scale-free network, an attack that targets the hubs can easily destroy such a network." — Barabási, Network Science (p. 318)

Random failures overwhelmingly hit peripheral nodes, so the system shrugs. Targeted hub removal fragments the same system after a handful of nodes. The two resiliences are close to anti-correlated — demonstrating one tells you nothing about the other.

The agent translation is direct. A fleet that has absorbed months of random failures — a flaky tool, an occasional timeout, a bad retrieval — has demonstrated random-failure resilience only. If your topology has hubs (the one orchestrator every workflow routes through, the single tool 60% of agents call, the retrieval index everything depends on), you are simultaneously fragile in a way that record does not measure. This is the same structural blindness that persistent homology surfaces from the other direction: the failure lives in the shape of the system, not in any node's health metric.

So the interventions that work are tail-targeted, not average-targeted:

• Cap the tail rather than optimize the mean. Per-trace token ceilings, step limits, and wall-clock budgets bound the distribution where the damage is. Mean cost per task falls out as a consequence.
• Route by predicted cost, not average cost. Length-aware routing cuts P99 by 25–69%; tuning for average throughput moves a number nobody experiences.
• Find the hubs and give them redundancy. Uniform redundancy is random-failure thinking. Redundancy concentrated at high-degree nodes is attack-tolerance thinking.
• Report worst-case alongside mean in evals. If task difficulty is heavy-tailed, a model can improve mean accuracy while making tail errors worse. Eval suites that report only aggregates cannot see that trade.

The five-question diagnostic

Before trusting any average in your agent telemetry, run this. It comes from the underlying statistics literature, and each answer is observable in data you already have.

Answer "Extremistan" to questions 2 or 3 and "asymmetric toward ruin" to question 5, and the Gaussian model is structurally wrong for your purpose. Not imprecise. Wrong about which regime you are in, and therefore wrong about which lever to pull.

Where the average is still right

The honest resolution is domain-dependent, and overcorrecting is its own error. Treating every distribution as fat-tailed without evidence sacrifices real analytical tractability for speculative tail modeling.

The distinction is not "averages bad." It is that the choice of distributional assumption is a modeling decision that determines which interventions are available to you, and most teams make it by default rather than by diagnosis. A team reporting mean cost per task has implicitly assumed finite variance. Very few have checked.

The takeaway

The mean is not a neutral summary. It is a claim that the distribution has a stable center and a bounded tail. For agent fleets — multiplicative chains, unbounded cost per run, hub-structured topologies, 35x median-to-P99 gaps — that claim is false, and every downstream decision inherits the error.

Pull last month's traces. Sort by cost, descending. If the top 1% accounts for more than the bottom half, stop optimizing the average and start bounding the tail. The runs that will hurt you are already in the data; they are just averaged out of the dashboard.

See also: Agent Economics for the unit-economics frame these distributions distort, The Agent Autopsy for what tail events cost in practice, Agent Observability for the telemetry this analysis consumes, and Persistent Homology for AI Agent Graphs for the structural view of the same blindness.