Where to Go Next

Jr Codex Agentic AI Notes

Level: All levels Prerequisites: Chapter 2: Building the Agent Time to complete: ~15 minutes

What This Curriculum Covered
The Six Ideas Worth Keeping
The Complete Jr Codex Arc
Staying Current
Directions for Depth
A Final Word

1. What This Curriculum Covered

The Arc
─────────────────────────────────────────
  Module 1   what "agentic" means, the spectrum, the
             reference architecture

  Module 2   the reasoning core — what it does well,
             what it does badly, tool design, MCP

  Module 3   memory within a run and across runs

  Module 4   planning, reflection, and the
             pathologies planning introduces

  Module 5   the frameworks, and when not to use one

  Module 6   workflows, durable state, multi-agent
             architectures and their failure modes

  Module 7   evaluating trajectories, tracing,
             debugging, recovery

  Module 8   injection, the lethal trifecta, human
             oversight, containment, governance

  Module 9   all of it, in one system
─────────────────────────────────────────

2. The Six Ideas Worth Keeping

1. USE THE LEAST AGENTIC SYSTEM THAT WORKS
─────────────────────────────────────────
  Most problems solved with agents are workflows.
  Autonomy is a cost you pay for flexibility you may
  not need.

2. CONTEXT IS THE AGENT
─────────────────────────────────────────
  The model reconstructs its understanding from the
  transcript every step. Agent quality is a property
  of what you assembled, not of the model you chose.

3. TOOLS MATTER MORE THAN PROMPTS
─────────────────────────────────────────
  Descriptions are the only information the model
  has about its options; error messages are prompts
  injected at the decision moment.

4. THE ORCHESTRATOR SEES WHAT THE AGENT CANNOT
─────────────────────────────────────────
  Loops, thrash, cycles and premature completion are
  invisible from inside. Detection belongs in
  deterministic code, always.

5. PREFER MECHANICAL VERIFICATION
─────────────────────────────────────────
  A compiler, a schema, a resolvable citation beats
  any amount of self-critique. Where you can choose
  the output format, choose a checkable one.

6. CONTAIN, DO NOT TRUST
─────────────────────────────────────────
  Assume the agent will attempt the worst thing its
  tools permit. Controls live in the executor, the
  credentials and the environment — never only in
  the prompt.
─────────────────────────────────────────

3. The Complete Jr Codex Arc

The Full Path
─────────────────────────────────────────
  Python              the tool
      ↓
  Data Science        the data discipline
      ↓
  Machine Learning    learning from data
      ↓
  Artificial          search, logic, agents, ethics
    Intelligence
      ↓
  Deep Learning       the architectures
      ↓
  NLP & LLM           language and large models
      ↓
  Generative AI       producing content
      ↓
  Agentic AI          taking action  ── YOU ARE HERE
─────────────────────────────────────────

What Recurred at Every Layer
─────────────────────────────────────────
  EVALUATION DISCIPLINE
    train/test splits ── benchmarks ── FID and
    CLIPScore ── trajectory metrics. The same idea:
    measure against something you did not optimise
    against.

  THE BIAS-VARIANCE INSTINCT
    generalisation in ML, overfitting a LoRA, an
    agent that memorises one path. Fitting the
    example rather than the pattern.

  COMPRESSION CREATES MEANING
    embeddings, latent diffusion, agent memory,
    handoff summaries. Forcing information through
    a narrow channel is what makes the channel
    meaningful.

  VERIFICATION DEFINES THE SYSTEM
    a test set, a schema, a citation, an approval
    gate. What catches the error determines what you
    can safely build.
─────────────────────────────────────────

4. Staying Current

The Filter
─────────────────────────────────────────
  IGNORE   framework release notes, agent
           benchmark leaderboards, "autonomous
           agent" demos, new orchestration
           libraries

  NOTICE   a new CAPABILITY (something previously
           impossible), a new PROTOCOL or standard
           with real adoption, a large price change,
           regulatory developments, and new ATTACK
           classes

  The last one matters most and is covered least.
  Agent security is where the genuinely new
  information appears.
─────────────────────────────────────────

A Sustainable Routine
─────────────────────────────────────────
  MONTHLY     re-run your eval set against current
              models. It is the only way to know
              whether to switch.

  QUARTERLY   re-audit the trifecta and the grants.
              Tool sets accumulate quietly, and
              yesterday's safe agent grew a
              capability last sprint.

  ONGOING     read incident write-ups, not launch
              announcements. Other people's
              production failures are the most
              useful reading in this field.
─────────────────────────────────────────

5. Directions for Depth

If You WANT TO SHIP AGENTS
─────────────────────────────────────────
  → Modules 3, 7 and 8 in depth. Memory, evaluation
    and control are where production agents are won
    or lost. Framework knowledge is not the
    bottleneck.

If YOU WANT AGENT SECURITY
─────────────────────────────────────────
  → Module 8 and the OWASP guidance for LLM
    applications. Build the injected-page eval case
    from Module 9 and try to defeat your own
    controls. This specialism is badly
    under-supplied.

If YOU WANT RESEARCH
─────────────────────────────────────────
  → the open problems this curriculum flagged:
    long-horizon planning past ~10 steps, calibrated
    uncertainty, multi-agent systems that reliably
    beat single agents, and injection defence that
    does not rely on containment alone. None is
    close to solved.

If YOU WANT BREADTH
─────────────────────────────────────────
  → back to the AI Notes' classical planning and
    multi-agent chapters, and the ML Notes'
    reinforcement learning. Agentic AI rediscovered
    many of those ideas; reading them properly makes
    the current work legible.
─────────────────────────────────────────

6. A Final Word

You have now completed the entire Jr Codex technical curriculum — from a first Python function to systems that plan, remember, use tools, and act under supervision.

Every layer built on the one before it. The Transformer you derived in Deep Learning became the language model in NLP, the text encoder in Generative AI, and the reasoning core here. Evaluation discipline from Machine Learning became trajectory scoring. The agent framework from the AI Notes became Module 1's architecture. Nothing in this final curriculum needed to be learned cold, which was the point of the order.

The specific tools will date. Frameworks will be replaced, model names will change, and some of the practices here will look quaint within a few years. What will not date is the judgement: knowing when a problem needs an agent and when it needs a function, knowing that context is the thing you actually control, knowing that the verification layer is the product, and knowing that a system which acts in the world must be contained rather than trusted.

That judgement is what makes someone worth hiring, and it is what this curriculum was actually for.

Jr Codex — 1-on-1 Personalized Coaching | Back to Module Index | Back to Agentic AI Index