Human coders reportedly show super-linear scaling on 14-day tasks

A researcher sharing a paper from their experiment comparing expert human coders with coding agents on the same 14-day tasks reports super-linear scaling for humans across different problems and agent setups. They interpret this as evidence of learning during problem-solving, rather than trying ideas without memory.

The researcher also recommends Claude Code session counts for different token budgets: one session for 5 million tokens; two independent sessions of 15 million each for 30 million; or three independent sessions for 100 million. With multiple sessions, the recommendation is to pick the best result.

They expect communication between agents to outperform these independent runs, but say they know of no controlled measurement of that comparison.