[pbjhsu]

The exhaustion people describe here resonates. But I think it points to something bigger: agents are great at tasks with clear success criteria (compile, test, deploy) and bad at tasks that require human judgment.

I work on a social platform and the frontier we're hitting isn't productivity, it's trust. Can an agent help two strangers build enough trust to have a real conversation? That requires progressive permission escalation, not task completion.

The interesting design constraint: the agent should never be able to reveal your identity to someone without your explicit approval. Unlike coding agents where more autonomy = more productivity, social agents need less autonomy and more consent checkpoints. Almost the opposite of what this thread is debating.