Why agents fail silently

Theory . The most insidious property of agent failure: an agent can return a
successful status while producing output that is quietly wrong — code that compiles,
passes partial tests, and still ships a bug. Students learn why distributed tracing is
the only reliable way to recover the missing context of which tools were called, in
what order, and where reasoning went wrong.

Use cases . The "green build, broken behaviour" case; an agent that reports
"done" after editing the wrong file; partial-test-pass masking a logic error — each
contrasted with what the trace would have revealed.

Practical exercises .

• Concept-check: given a "success" status plus its trace, find the discrepancy that
  proves the run silently failed.
• Applied: design a check that would have flipped a specific silent failure into a
  loud one (an assertion, an extra test, or a trace-based alert).