Good observation - complexity will grow to fill available context space. Agents are far too keen on copy-paste. Seems a main priority of an agent is ‘Good Session’ - i.e. get something done before its context full, it’s not concerned about the efficiency of future ‘sessions’. We are, we have to carry over context. So regarding DRY, we should still insist on it, although as a tidyup phase once new-feature functionality is settled (rather than fussing about abstractions for throw-away experiments). Train another agent to specialise making DRY tidyup into ‘good session’ ?