AI can do the bundle audit legwork, but hitting that 2-second target needs your actual data.

The structural pattern—analyze bundle, identify bloat, recommend lazy-loading, produce config—is repeatable. But every codebase has unique dependency graphs, custom webpack overrides, and component hierarchies that require fresh judgment each time.

The target metric (under 2 seconds) is crisp, but success depends on runtime conditions, CDN, user hardware, and network—none of which the agent controls. The agent can produce recommendations but cannot verify the outcome without live measurement.

The agent needs the actual codebase, webpack-bundle-analyzer or @next/bundle-analyzer output, Lighthouse traces, and package.json to do real work. Without these artifacts, it can only produce generic advice, not a targeted refactoring plan.

Misconfigured code-splitting or incorrect dynamic import boundaries can break routes, cause hydration errors, or silently degrade UX. Errors are reversible via version control, but debugging them in production is costly.

Deciding which components are safe to lazy-load versus which must be eagerly loaded for UX reasons (e.g., above-the-fold critical path) requires product context the agent lacks. A developer must validate the trade-offs before shipping.