Building Agents: A 3 Year History (www.dallinbentley.com)

This piece is a three‑year retrospective on building AI agents, tracing the arc from first contact with ChatGPT (Nov 2022) through the frantic experimentation of 2023, workflow engineering in 2024, and the agent era of 2025 — when recursive LLM loops, better reasoning models (e.g., OpenAI’s o1), and multi‑agent systems became product‑viable. It explains why early limits — small context windows (text‑davinci-003’s 4,096 tokens), unreliable one‑shot outputs, and brittle prompt engineering — drove adoption of retrieval‑augmented generation (RAG) and vector DBs, and why patterns like DAG/workflow decomposition, chain‑of‑thought prompting, and engineering hacks (Cursor’s “fast apply” fine‑tune) mattered for reliability. Technically, the writeup highlights the shift from encoding rigid stepwise workflows to building flexible environments and toolkits that let agents act — including allowing agents to generate code (Manus AI style) to call APIs — and the move toward combined “Analyst” agents that can explore data, run SQL, then visualize. It also identifies remaining research and product challenges going into 2026: encoding heuristics for vertical workflows, selective and usable long‑term memory, and flexible autonomy/triggering. The significance for AI/ML practitioners is clear: success now depends less on prompt tricks and more on designing tools, environments, and codified workflows that enable robust, multi‑step agent behavior.