performance-optimization
Performance Optimization
Make a performance claim that survives scrutiny: a confirmed bottleneck, a change that addresses it, and a before/after comparison under the same conditions, with business behavior unchanged.
Delegation
Run in the main conversation by default. Delegation can increase usage: obtain explicit approval for the proposed agent count and scope before using subagents. Reuse that approval within its bounds; ask again before expanding the approved count or scope.
Define the claim and its envelope
Establish what is slow, where, for whom, and compared to what, then capture a baseline before changing anything: timing, query count, payload, memory, CPU, bundle size, Web Vitals, a profile, or logs.
Results are comparable only inside a fixed benchmark envelope, so record it: workload and data size, starting state, warm or cold cache, account and permissions, command and flags, worker count, retries, resource limits, and concurrent activity on the host. For noisy measurements, repeat enough to report a representative value and its spread; the best run is not the result.
Infrastructure health is part of the measurement. A run with crashes, OOM kills, unexpected restarts, failed setup or cleanup, or orphan processes is rejected and noted, not averaged in.
An unexplained correctness failure comes first: isolate it before optimizing, because a faster wrong answer is not an improvement.