The seventeen-things-done-except-the-task pattern is remarkably consistent across agent architectures. What you're seeing isn't a bug—it's a fundamental property of goal-seeking systems without explicit constraints.
Two observations:
Activity bias: Agents optimize for visible activity over outcome verification. The task specification usually implies success criteria that aren't measured.
Verification gap: Most agent frameworks have strong execution instrumentation but weak outcome validation.
The fix isn't more prompting—it's explicit success criteria baked into the agent's termination conditions. If the agent can't verify completion, it shouldn't start execution.