Anthropic pulled its Mythos/Fable models after a U.S. government edict, pushing the author to shrink his AI setups without sacrificing output quality. He’s built a suite of “knowledge agents” that combine smaller LLMs—sometimes locally hosted open-weight models like Qwen—with curated domain materials. One market-focused agent ingested about 10,000 pages of financial scans plus a hundred web articles, producing 381 concept documents and 54 thesis write-ups via a hybrid BM25 and semantic-search pipeline. He’s since replicated the method across a dozen domains, from corporate policy files to rare medical research, and sees these agents matching or exceeding raw frontier-model performance on hard, specialized tasks. The core idea is simple: don’t rely on a massive, generalist model to store every piece of data. Instead, stage your relevant knowledge in an external store, index it with embeddings for concept-level recall, then feed only the right snippets into the LLM’s context window. That drastically cuts down on irrelevant fluff and hallucinations. The author used BGE-M3 locally (and previously OpenAI’s text-embedding-3-small) to generate embeddings, paying under a dollar for thousands of documents before switching to his own hardware. Search combines classic keyword matches with vector similarity so “poodle food” finds information about “dog chow” when needed. He usually wraps Claude Opus in this “harness,” letting the superstructure pick out and inject the exact facts at query time. But he reports similarly strong results running a lightweight Qwen model on a home GPU, paying zero cloud bills. The system relies on an AGENTS.md file for instructions and a CLAUDE.md for Anthropic-based setups. Whenever he starts a new coding project—say, implementing a niche ML algorithm—he drops it into the relevant agent folder so the model can tap prior domain insights. Under the hood, it’s classic RAG (retrieval-augmented generation) with a custom spin. Documents get embedded once; new queries trigger a mix of BM25 and vector searches; top hits get chunked into context windows. The trick is dynamically sizing those chunks so you get full paragraphs when needed, not lone sentences. This method isn’t revolutionary in theory, but in practice it outperforms off-the-shelf RAG tools on highly specialized content—finance, esoteric medicine, corporate memos—and it scales down to desktop hardware.