Lots of products working on saving-tokens-space. Compression, Tool Output rewrite, Sitting as proxy Cache between harness and provider , doing circus with interceptor hooks - these are some of the approaches we are seeing today.
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) ?
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) ?