Assume everything fails

Adopt this stance up front. In a single agent turn, any of these can go wrong:

• The model API returns an error, times out, or rate-limits.
• The model returns malformed tool arguments, or asks for a tool that doesn't exist.
• A tool throws (an external API is down, a file is corrupt).
• The stream connection drops mid-turn.
• A downstream store (database, object storage) is briefly unavailable.

A naive agent crashes the whole turn on any of these. A reliable one has a defined response to each. Reliability is the sum of these defined responses.