Scaling Anti-Patterns

Seven scaling anti-patterns that we have either committed or observed — and the damage each one causes.

The Big Bang: deploying 5 agents simultaneously instead of one at a time. Coordination efficiency dropped from 94.7% to 78.3% in one day because 5 new agents created 25+ new interfaces simultaneously. Recovery took 2 weeks. The Phantom Benefit: adding an agent because the team "might need" the capability, without modeling demand. The agent sits at 15% utilization while consuming coordination overhead at 100%. The Flat Scaling: adding agents without restructuring the coordination topology. Going from 12 to 20 agents in a flat hierarchy turned CLAWMANDER into a bottleneck processing 200+ messages per day.

The Premature Optimization: splitting roles before the volume justifies it. Creating two agents where one sufficed adds coordination overhead that exceeds the throughput gain. The Clone Army: adding multiple agents with the same role instead of scaling the existing agent's capacity. Clones compete for work, produce inconsistent outputs, and create boundary conflicts. The Scale-and-Forget: deploying a new agent and not recalibrating baselines, leading to false alerts for weeks. The Token Bonfire: scaling to 20+ agents on premium models when 12 agents on standard models would produce equivalent effective throughput at 40% of the cost.