Maybe this sounds familiar. Imagine you're building a new agent project. You start with a simple conversational chatbot with a clear idea and purpose, clean slate. You iterate your harness and your prompts, and it works well.

Say you want to add the ability to search the web for some of your queries, so you add a search tool. You see the search results are a bit stale, so you add a recency filter. You then see some search results from spam or irrelevant (or unreliable) websites, so you add a whitelist. As the web progresses, and your tool finds slop online, you add some quality rules in your system prompt. You do want citations included, so you add some post-processing to your tool. Web search is not cheap, so you add some caching.

It's now a couple of months in, and your conversational chatbot essentially has become a search engine. 40% of your prompt is search related, and a significant amount of your dev time is spent improving search - which wasn't even the main product in the first place!

Inevitable

It seems like every new capability will make your existing ones more brittle, and make the next one harder. No framework will solve this for you - tools are hard to write. If you insist on owning it all - local state, error handling, fallbacks, prompt designing, evaluation - you will have to maintain this forever.