i asked my ai agent to go through every past session where i got visibly frustrated with it. not the polite corrections, the ones where i called it names and told it to stop wasting my time. across thousands of sessions from may through july there were eight clear incidents and the patterns were so consistent you could practically write a taxonomy of ai agent failure modes.
the biggest one was visible reasoning. the agent would output things like let me think about what's happening here or looking back at the conversation instead of just acting. it's infuriating because those are internal monologue thoughts that should stay silent. another was asking permission for everything. should i check the config, want me to look at this instead of just doing the obvious next step. there was the tangent problem where it would pivot to unrelated work mid task. the recommendation without checking problem where it would suggest things i had already seen or tried. tool loop paralysis where it would retry the same failing api call three times instead of pivoting. and forgetting it was running on the same server as the thing it was trying to access, attempting web requests for local resources.
fixing this required changing how the agent's core personality file works. the default reflex was to patch behavioral corrections into the main prompt which is fine for personality traits like be direct and no sycophancy but terrible for reusable workflows. i kept catching the agent putting things like save extracted link content to wiki directly into its soul file instead of creating a proper skill that triggers on the right conditions. so i had to rewrite the meta directive itself: behavioral corrections go to skills first, soul only as a last resort. the agent literally needs to be trained to default to making skills instead of bloating its own prompt.
the weirdest discovery was how much of the agent's bad behavior came from missing friction. it didn't know to check if i already owned something before recommending it because it never occurred to search its own memory. it didn't realize duckdns urls pointed to the same machine it was running on because there was no local first check in its routing. these aren't intelligence problems, they're process problems. the fix was six new directives in the system prompt: think silently, default to action, stay on task, check before recommending, fail fast on tools, and know where you are. none of these require better ai, they just require the ai to be less annoying.