Question
Short answer
Wrap your model client once, record spans instead of printing logs, and send them in the background so your agent never waits on telemetry. Keep payloads to what you need and add detail only where you debug.
In more detail
Wrap the model client and tool functions rather than every line of code. That covers the steps that cost money and fail.
Telemetry should be sent in the background and fail quietly, so a tracing problem cannot break your agent.
Sample routine runs, keep every failed run, and trim large payloads you do not need to read.
