CJ Avilla from Anthropic sent his Claude Platform OpenTelemetry data to Sentry. It turned up an old project that was still burning tokens.
Every retry costs tokens, and the cause can be a slow API or database, not the model. In this live workshop, you’ll see:
A Claude agent’s failed tool calls, retries, and final success in one trace
Whether the fix belongs in the prompt, the tool, or the service behind it
Source: ByteByteGo (system design interviews) — Read the full article on the original site ↗