The way coding agent work is fantastically wasteful. All the megabytes of code are processed over and over and over, sometimes withing just one session.
There are papers describing KV cache precomputation for commonly used documents (e.g. KVLink), but, of course, it's not a priority for model providers: they'd rather sell you more tokens, also they would rather get to AGI/ASI first than optimize usage of existing models...
There are papers describing KV cache precomputation for commonly used documents (e.g. KVLink), but, of course, it's not a priority for model providers: they'd rather sell you more tokens, also they would rather get to AGI/ASI first than optimize usage of existing models...