The agent adds a new settings page and wraps the whole file in a brand-new <ErrorBoundary fallback={<p>Something went wrong</p>}>. The app already has a route-level boundary, a feature-level one around the editor, and a shared fallback that reports to the logger. The new wrapper swallows the error, skips the report, and looks different from every other failure screen.

I’ve started treating a one-off error boundary in an agent PR as a smell. If the repo already has a map — which routes own a boundary, which widgets are isolated, what the fallback is — the model should use that map. It should not invent a fourth kind of “something went wrong.”

Why a new boundary is usually wrong

The model has seen a million examples that start with “wrap it in an error boundary.” That advice is correct at the architecture level and wrong at the leaf. Two boundaries in a row, with different fallbacks, make recovery harder to reason about. A boundary that does not call the existing logError helper makes the page look fine in the UI and silent in the dashboard.

There are real reasons to add a boundary: a third-party widget, an isolated editor, a route the team has not wrapped yet. Those reasons should be in the PR, and the fallback should be the shared one unless you are designing a new tier.

A simple boundary policy

  • Read the existing map first — app/error.tsx, src/lib/boundaries.md, the router wrappers
  • Reuse the shared fallback and logger
  • Add a new boundary only for a new isolation tier, named and documented on the map
  • No inline fallback={<div>Error</div>} in a random page

If the framework already provides route error elements (Next.js error.tsx, Remix ErrorBoundary), prefer those over a custom wrapper in the page module.

What to put in the prompt

Here is how errors are isolated in this app (see src/lib/error-boundaries.md or the root ErrorBoundary). Do not add a new ErrorBoundary around a page that is already covered. If this feature needs isolation, use the shared fallback and logger, and update the map. Do not invent a one-off “Something went wrong” message. 

Paste the map or the existing boundary component. In review, search the diff for ErrorBoundary and componentDidCatch. New hits need a reason.

Async errors are a special case. A render boundary will not catch a failed fetch unless you map that failure into state or a router error. If the agent added a boundary “to handle the API error,” it probably did not. Point it at the existing query-library error UI or the route’s error.tsx, and keep the boundary for render-time failures.

If the new page is a modal or a drawer, isolating it can be right — a crash in a dialog should not blank the whole settings route. Use the shared fallback, still, and say on the map that the dialog is an isolation tier.

Reject / send back

Reject

  • A page-level wrapper that duplicates the route boundary
  • A fallback that does not report to the existing logger
  • A unique error illustration for one page, unless that is an explicit request
  • Swallowing the error and rendering an empty state

Accept

  • A feature that relies on the existing route or app boundary
  • A new, documented isolation tier with the shared fallback
  • A framework error.tsx that matches the rest of the app