Why an agent must be observable

Agents are non-deterministic: same input, different paths.

So — if you can't see each step, you can't debug. Observability should show at least:

• each model decision (which tool it wants, what it says);
• each tool call — arguments and result;
• where time is spent or failures happen.

Without it, you're guessing.

One line:

Debugging an agent is really about reading its trace.