HACH team designs AI agent to keep working when one tool fails

The design note draws a clear line between two failure modes that are easy to conflate. When the LLM provider itself goes down, the agent cannot generate any response, so waiting messages and queuing are the only options. But when just one integration—a calendar API timing out, a CRM losing authentication—is unhealthy, the LLM is still working, and stopping the whole agent discards information it could have delivered. The note identifies this as the most easily overlooked failure pattern.

The proposed patterns build on that distinction. Morning reports are assembled from independent fetch results rather than a single all-or-nothing batch, and the note points to Promise.allSettled as the mechanism that keeps one source's failure from dragging down the others. Separately, the design insists that 'zero items' and 'failed to retrieve' never share the same wording, because a user cannot tell them apart and may go days without noticing a broken connection. For integrations that repeatedly time out, the note suggests temporarily skipping the fetch with a note based on the last successful data, then using a lightweight health check to detect recovery and return to the normal path automatically.