Amey-Thakur Agent capacity planning
Overcommitment is the most reliable way to slow a company down: every item in progress, nothing complete, and quality falling everywhere. Capacity planning is the discipline of knowing the real ceiling and treating it as a constraint rather than a suggestion.
Method
- Measure capacity from what actually completed. Last cycles' throughput is a better estimate than anyone's optimism, and it already accounts for the interruptions nobody plans for.
- Count both budgets. Agent runs cost tokens and money; humans cost hours and attention. A plan that fits one and not the other does not fit.
- Reserve for the unplanned. Incidents, support, and urgent requests consume a predictable share, and a plan at full capacity fails on the first of them.
- Limit work in progress explicitly. Fewer items finished serially beats many in parallel, because value arrives at completion rather than at start (see agent-work-assignment).
- Make the trade visible when refusing. Taking this on means dropping that, presented as a choice for the human rather than a refusal.
- Re-plan when reality diverges, not at the cycle end. A plan already known to be wrong is worse than no plan, because it is still directing work.
- Track estimate against actual. The gap is what improves the next estimate, and it is usually consistent enough to correct for.
Boundaries
Planning allocates capacity; it cannot create it, and a persistently oversubscribed company needs scope cut or resources added by a human decision. Agent throughput estimates are rough and change with model and task. Human capacity involves people's lives and is not a resource to optimise without their say (see agent-people-ops-desk).