In my tests, honest conclusion from repeated observations is that the tangible results are possible only by reducing Turns upfront in an agentic loop. Turns saved in first 25% of an otherwise 40 turn loop can produce significant gains - than turns saved in the latter 50% of the same loop. I don't think I need to explain it here - the compounding context.
From that observation, shouldn't there be a lot more energy put into tool engineering (that all other xyz engineering) ?
2 comments