An agent is not a smarter chatbot. It is a stack.

From Models to Agents maps the eras: chat, reasoning, tools, then the enterprise approach. AI Right Now Part 6 covers control: what an agent may touch and what you must approve. This series is the wiring diagram: eight parts that work together as one machine.

If you only remember the picture, remember this: the language model is an engine. Memory, planning, tools, retrieval, and a loop around them are what make it an agent. Evaluation is how you know the machine did what you think it did.

The stack

Goal
  │
  ▼
Orchestration  (who calls what, when to stop)
  │
  ├── LLM               language engine
  ├── Memory            what persists across steps
  ├── Planning          goal → steps → re-plan
  ├── Function calling  structured tool requests
  ├── Tools             the runtime actually executes
  ├── RAG               retrieve, then ground
  └── Evaluation        score the path, not the end result

Read the diagram top-down once, then left-to-right as a checklist. Each part has a simple picture to help you remember it. Orchestration runs the loop, like a clock. The LLM proposes the next step. Memory stores what happened, like a notebook. Planning keeps the list of steps, like an agenda. Function calling is how the model makes a request, like an order form. Tools are the actions the agent can take. RAG fetches approved documents, like a library card. Evaluation checks the whole run, not only the final answer, like an inspector.

None of these pieces is optional in a serious system. You can hide them behind a vendor brand. You cannot delete them.

Three named gears

These are not eight marketing terms. They are a few well-documented parts, connected together.

The loop. ReAct (external link) (Yao et al.) interleaves a thought, an action, and an observation from the environment, then continues. Google’s short note (external link) is the public explainer. In this series that loop is planning plus tools plus memory: the agenda changes because the world talked back.

The knowledge. Lewis et al. (2020) (external link) named retrieval-augmented generation: fetch from a corpus you index, then condition the next tokens on those passages. That is RAG. Acting after the quote is a later choice, not a different kind of model.

The runtime. Someone has to call the model, run the tool, append the observation, and stop. OpenAI’s agents guide (external link) is one instance of that scheduler. A while with a max-step counter is the same architecture in a generic term.

The eight parts unpack those gears. They do not replace them.

The series

  1. The language model — the engine is not the vehicle
  2. Memory — working context, session, longer-lived store
  3. Planning — goal → steps, then re-plan after observations
  4. Tools — callable capabilities the runtime runs
  5. Function calling — name + arguments; model proposes, code executes
  6. RAG — retrieve from an approved corpus, then ground
  7. Orchestration — the loop and the stop
  8. Evaluation — measure the trajectory

Sources