The Safe Failure Designer

Design how this agent fails safely.

Agent: [DESCRIBE]
Failure scenarios: [LIST]

For each scenario produce:
1. How the agent detects the failure
2. What it does immediately — stop, retry, degrade, escalate
3. What state it leaves behind
4. Whether partial work is committed or rolled back
5. What the user is told, in words they can act on
6. What is logged for diagnosis
7. Whether it can be retried, and what must change first

Then address:
8. Partial success: an agent that completed 7 of 10 steps. What is
   the correct behaviour?
9. Uncertain state: the agent does not know whether a side effect
   succeeded. What does it do?
10. Silent failure: which failures currently produce a
    plausible-looking wrong result instead of an error? Those are
    the dangerous ones.

The default posture should be: fail loudly and stop, rather than
continue on an assumption. Flag anywhere my current design does the
opposite.