gave an AI agent 200K tools to see when it gets confused, upto 10k: fine at 200k: success 85% → 30%, it usually finds the right tool, then confidently picks its evil twin (daybay_v2 instead of daybay), neither a smarter search fixed it nor a faster one.
it's just k-onfused!!
built with @claudeai Haiku 4.5 via Claude Code, 200k tools in @qdrant_engine, embeddings + re-ranking run locally with @huggingface transformers.js.
wanted to try @typesafeai's Jev for picking between the twins. next run?? 👀
Blog: https://t.co/IsWFmc9h43
Repo: https://t.co/0wzV32Ar8l
was sure my p2p catalog sync needed a consensus algorithm. then realized entries only ever get added, so merging two catalogs is just a set union. no consensus needed.
apparently this has a name (CRDTs, grow-only set)