A_ made by arjun

AI Agent learning path

From Transformers to Memory: How an AI Agent Works

These five articles do not begin with a framework API. They follow Context, Runtime, Feedback, Planning, and Memory to build a working picture of an Agent system.

Start with the first article
  1. 01 Context
  2. 02 Runtime
  3. 03 Feedback
  4. 04 Planning
  5. 05 Memory

This is a learning order, not a fixed component call graph.

Five articles, one thread

Follow the problems the system must solve

Each article adds one layer. The next article keeps using the concepts established before it.
  1. 01
    Context

    How Transformer Self-Attention Builds Context

    Start with how one token gathers information from other positions and becomes a contextual representation.

    Question answered How does a model decide which parts of a sequence matter to the current token?

  2. 02
    Runtime

    How Does an LLM Agent Actually Finish a Task?

    Treat an Agent as an execution system: the model decides, tools change the outside world, and state keeps track of progress.

    Question answered How does one model call become a loop that can keep moving a task forward?

  3. 03
    Feedback

    How ReAct Agents Act, Observe, and Adjust

    See why Action and Observation belong together, and how real feedback changes the next decision.

    Question answered How does an Agent revise its judgment after a tool returns a real result?

  4. 04
    Planning

    How Agents Plan, Search, and Replan

    Separate task decomposition, search, and replanning so each method has a clear job.

    Question answered When should an Agent plan first, and when should it replace the plan?

  5. 05
    Memory

    When an Agent Forgets, Look Beyond the Model

    Separate Context, State, and Long-Term Memory, then follow the write and retrieval paths through the system.

    Question answered What should be retained, and how does it return to context when needed?

After the path

You will have a system map you can keep extending

The goal is not to memorize framework names. It is to locate the likely layer when an Agent misjudges, fails a tool call, drifts from a plan, or forgets information.

  • Separate model context from application-managed state.
  • See how the Agent loop and ReAct feedback loop relate.
  • Understand why Planning and Memory belong to system design.
Begin the path