AI Agents from First Principles (goyalpramod.github.io)

This piece walks through building AI agents “from first principles,” stripping away heavyweight frameworks to show how to assemble reliable, purpose-built agents using plain Python and core libraries. The author argues that current agents (early 2025) are unstable and often hallucinate, so developers should construct agents from modular building blocks—clear prompts, a chosen model, callable tools, memory, and Retrieval-Augmented Generation (RAG)—rather than rely on opaque abstractions. The guide is practical and opinionated: favor explicit prompts with roles, examples and tagging; treat the LLM as the CPU and choose between API providers (OpenAI/Anthropic/Google) for speed and convenience or self-hosting (Llama/Qwen/etc.) for cost and control; and use deployment techniques (quantization, batching, vLLM) when self-hosting larger models. Technically the post gives runnable patterns: a simple @tool decorator that exposes function metadata (is_tool, description, parameters) for schema-based tool calling; memory split between context window (short-term, limited by model windows — e.g., Claude ~200k tokens, GPT‑4 ~128k tokens) and long-term stores with embedding-backed retrieval; and a RAG pipeline using SentenceTransformer embeddings plus vector DBs (Qdrant, Weaviate, Pinecone, Chroma, Milvus). It also covers document chunking and retrieval strategies, emphasizing that retrieval + relevant context beats naive fine-tuning for keeping agents up-to-date and consistent over long sessions.