Coding agents are fun until you realise the real boss is not the model, it is the stupid context

We were building this coding agent for Cheetah AI , and everything looked cool in the demo world, then real repos hit and boom, welcome to “the file got corrupted, let me rewrite the whole file ” hell because all context window got used up.

So the only way out everyone recommends is “just use rag and index the whole codebase (thanks to augment code blogs for motivation that was never needed)”

Yeah sure, man, let me just burn my gpu and my wallet every time someone does git commit -m “fix typo”

indexing a big repo, waiting, reindexing on every change, watching tokens vanish faster than motivation on a Monday, all so the model can still miss the one file that actually matters
at some point i snapped and thought what if we just stop pretending embeddings are magic and literally grab what we need from the code like normal people

So Greb grebmcp. com was born
no fancy “ai native yada yada”, it just does ruthless grep, glob, read, ast parsing, tears the repo apart, pulls only the useful chunks, and then reranks it so the model gets exactly what it needs instead of a result

Small LLMS suddenly got a lot better than I thought, seriously. Now I am using Grok code on Copilot, and I literally thought I was using Claude Sonnet

Challenges were the usual
How do we make this fast enough that it does not feel like running yarn install
how do we stop the prompt from becoming a novel while still giving enough code so the agent does not hallucinate a whole microservice

but honestly the fun part is watching it beat setups that throw way more hardware and bigger models at the problem
just because we stopped reindexing the universe and started treating code like code instead of generic text