From chatbot to agent: how tools, planning and memory were added step by step

_jaydeepkarale · x · 2026-11-04

A narrative explainer on how AI agents were built step by step, starting from ChatGPT's launch:

Starting point: ChatGPT (Nov 2022) could talk — answer questions, summarize docs — but a chatbot can only explain why checkout is slow, not actually go check.
Adding tools: an engineer wired up metrics, logs, and database queries; the model picked a tool, read the result, decided next steps — loop that and you have your first agent.
Adding planning: agents looped forever or picked wrong tools and kept going, so it wrote steps first, verified each result, with a hard step limit.
Adding memory: everything vanished when the session ended ("a goldfish with a PhD"), so short-term memory for the task and long-term memory for users and facts were added.

A problem-driven progression that clearly explains where the three core agent components — tools, planning, memory — come from.

More from coding & agent