The Capacity Forecasting Agent: A Weekly Workflow

An agent that runs every Monday, forecasts the week, and files tickets for the bottlenecks. The forecasting model, the ticketing integration, and the false-alarm rate.

Weekly cadence

Weekly is the right cadence for most capacity planning. Every Monday at 9 AM the agent runs forecasts for traffic, storage, compute and files tickets for projected bottlenecks; daily is too noisy and monthly is too late. The fixed cadence builds the team’s habit so “capacity tickets show up Monday” becomes part of the routine.

• Monday 9 AM. Forecasts for the week; tickets for projected bottlenecks.
• Weekly is right cadence. Daily is too noisy; monthly is too late.
• Builds team habit. Capacity tickets show up Monday; part of the routine.
• Per-team rhythm. The cadence aligns with sprint planning; supports continued attention.

Forecasting model

The forecasting model has three properties. Time-series forecast on each metric (traffic, storage growth rate, CPU utilisation) with seasonal adjustment for weekly and monthly cycles; confidence intervals because the forecast is a band not a line and the width is part of the output; out-of-distribution flags when recent data deviates from the historical pattern.

• Time-series with seasonality. Traffic, storage growth, CPU; weekly and monthly cycle adjustment.
• Confidence intervals. Forecast is a band, not a line; band width is output.
• Out-of-distribution flags. Recent data deviates from historical pattern; forecast flagged uncertain.
• Per-forecast uncertainty captured. The output includes its own confidence; supports correct interpretation.

Ticket integration

Tickets are the action surface. When projected utilisation crosses a threshold (e.g., 80% of provisioned capacity within the week), file a ticket; the ticket includes the metric, the projection, the recommended action (scale up by N units), and the deadline; tickets close when the action is taken or the projection improves but the agent does not auto-close because humans verify.

• Threshold-based filing. 80% of provisioned capacity within the week; the trigger.
• Ticket contents. Metric, projection, recommended action, deadline.
• Close on action or improvement. Either the action is taken or the projection improves.
• No auto-close. Humans verify; the agent does not silently retire tickets.

False alarm rate

False alarms have a target band. Forecasts are noisy and some tickets close without action because the projection improved; the false-alarm rate target is under 30% (higher means the model is too sensitive, tune the threshold); lower than 10% is suspicious because the model is probably missing real bottlenecks (calibrate by checking what the team actually scaled in the last quarter).

• Target under 30%. Higher means the model is too sensitive; tune the threshold.
• Under 10% suspicious. Model probably missing real bottlenecks.
• Calibrate against reality. Check what the team actually scaled in the last quarter.
• Per-quarter calibration. The target band reviewed each quarter; supports continued accuracy.

Escalation paths

Three escalation paths cover the urgency spectrum. Standard ticket filed in the team’s queue; urgent ticket (projected breach within 48 hours) also notifies on-call; critical projection (already breached) pages immediately because a forecast that arrived too late is still useful.

• Standard ticket. Filed in the team’s queue; the routine path.
• Urgent ticket. Projected breach within 48 hours; also notifies on-call.
• Critical projection. Already breached; page immediately; the forecast still informs.
• Per-urgency routing. The routing mirrors urgency; supports correct response timing.