AI EngineeringZero to ProductionHome·About·Contact
Part 0 · The Path

Learning Roadmap & Common Mistakes

The step-by-step path from basics to your first real agent, mapped onto this course — plus the beginner mistakes that waste the most time, and how to dodge each one.

🧭 5-step path⚠️ 5 mistakes🏗️ Capstone plan

The roadmap

The classic beginner path — with the exact place in this course where each step happens.

  1. Learn the basics

    What agents are, the building blocks, the loop, LLM-vs-agent. Build the mental model before writing code.

    Ch 0 · Foundations (you're basically here)
  2. Learn to use the tools

    Make real API calls, control the model, get reliable structured output back. The raw materials of every agent.

    Ch 1 · First call + Ch 2 · Prompting & structured output
  3. Build a small project

    Ground the model in real data (RAG), then give it tools and a loop (agent). Start with one simple, working thing.

    Ch 3 · RAG + Ch 4 · Agent
  4. Test & improve

    Measure quality with evals so you can prove a change helped. This is how you earn the trust to remove the human.

    Ch 5 · Evaluation
  5. Ship & harden

    Caching, observability, guardrails, deployment — turn the prototype into a background service you can rely on.

    Ch 6 · Production hardening
The rhythm underneath the roadmapConsistency · Practice · Mastery. Short sessions where you actually run and break code beat long sessions of passive reading. Read a little, run a lot.

Common mistakes to avoid

1 · Starting too complex

Trying to build the fully-autonomous dream on day one.

✓ Fix: solve one simple slice first, then grow it. (Ch 0 scoping rules.)
2 · Vague / weak prompts

Fuzzy instructions → fuzzy, inconsistent behavior.

✓ Fix: the 5-part prompt + examples. (Ch 2.)
3 · Not testing outputs

Shipping on "seems fine," then quality silently drifts.

✓ Fix: an eval harness + regression gate. (Ch 5.)
4 · Skipping the basics

Jumping to agents without understanding one API call.

✓ Fix: do Ch 1–2 before Ch 4. The loop is just many calls.
5 · No human-in-the-loop

Letting the agent take irreversible actions unchecked.

✓ Fix: approval gates on risky steps. (Ch 4 & 6.)

The three mantras

Tape these above your desk
  1. Start simple. The smallest thing that works, then iterate.
  2. Test everything. If you can't measure it, you can't trust it.
  3. Keep a human in the loop until the evals say you don't need one.

The capstone — your AI DevOps Engineer

This course's destination is Part V: Ch 7 teaches the Forward Deployed Engineer method, and Ch 8 assembles everything you've built into one advanced project — an AI DevOps Engineer that onboards into any company (Git/GitLab, Terraform, Kubernetes/EKS, AWS, Docker, Jenkins, CI/CD, Argo CD, Python) and can even be hired on subscription. Same "start simple → test → increase autonomy" discipline, applied to infrastructure:

autonomy is earned one rung at a time — evals unlock each step 1 · Observe 2 · Recommend 3 · Act (gated) 4 · Selective reads diagnosis reviews every PR approves each action prod stays gated The autonomy ladder. Start read-only (Observe), then propose changes (Recommend), then execute reversible ops on approval (Act), then let a proven-safe allowlist run unattended in non-prod (Selective). Each rung unlocks only when evals prove the previous one safe — and prod stays human-approved.
🗺️ How to read this diagram

This staircase is the whole plan for the course's final project (the capstone), drawn as a ladder of trust. Read it left-to-right and bottom-to-top: each box is a stage where you let the AI agent do a little more on its own, and each step up is only unlocked once you've proven the step below is safe. You climb one rung at a time — you never jump straight to the top.

  • The four boxes are the rungs, and they get higher as you go right — that rising height is the picture's way of saying "more freedom, more responsibility." Rung 1 · Observe is the safest: the agent only looks and reports what's wrong (it "reads diagnosis"), changing nothing.
  • 2 · Recommend: the agent now suggests fixes but a human still checks each one — the label underneath says it "reviews every PR" (a PR is a proposed code change someone must approve before it takes effect).
  • 3 · Act (gated): the agent is allowed to actually do reversible things (like restarting something), but only after a human clicks approve — "approves each action." "Gated" means a gate/checkpoint stands in front of every action.
  • 4 · Selective: the top rung — a short, pre-approved list of safe actions runs on its own, but the red note "prod stays gated" warns that the real live production system still always needs a human. That red colour is a deliberate warning.
  • The short diagonal arrows between the boxes are the climb — you move up only in order, and the caption's rule tells you what powers each climb: "each rung unlocks only when evals prove the previous one safe." Evals are automated quality tests (you'll meet them in Ch 5); passing them is your evidence to earn the next rung.
  • Notice this mirrors the numbered roadmap higher up the page. The 5-step path (learn basics → learn the tools → build a small project → test & improve → ship & harden) is the same idea in words: start tiny and safe, measure, then grow. This ladder just applies it to how much you trust the agent.

In short: don't read this as "steps to full automation." It's a trust ladder — you spend evidence from tests to buy one more rung of freedom at a time, and the very top of production stays human-approved on purpose. Start at Observe, prove it, then climb.

PhaseGoalHuman roleUses
0 · PrepFinish this course. Brush up your DevOps tools (you already know most).All chapters
1 · ObserveAgent diagnoses incidents read-only across K8s/AWS/CI; answers "how do we do X here?"Reads the diagnosisCh 2, 3, 4
2 · RecommendAgent opens PRs/MRs: fixes, Terraform, Dockerfiles, pipeline repairs.Reviews every PRCh 4, 6
3 · Act w/ approvalExecutes gated reversible ops (scale, restart, sync/rollback) on one-click approval.Approves each actionCh 5 (evals gate it)
4 · Selective autonomyA short allowlist of proven-safe non-prod ops run unattended; prod stays gated.Exceptions onlyCh 6
🏗️ The single most useful question to answer firstWhich one painful, repetitive DevOps task should the thin vertical slice tackle? (e.g. "diagnose a crash-looping pod.") Nail one workflow end-to-end — read-only, safe — before adding any more. That's the FDE method from Ch 7.
Reality check on "fully autonomous"Phases 1–4 keep a human approving anything that changes state, and prod stays human-approved indefinitely. A lights-out agent running terraform apply on prod is not the goal — a copilot that turns an hour of toil into a 60-second PR review is. Autonomy is earned per-operation, on evidence from evals.
When you reach the capstoneCome back and ask me to help scope Slice #1 in detail — the exact read-only tools, the RAG sources (your runbooks), the sandbox (kind/minikube + a budget-capped AWS account), and the Diagnosis prompt. We'll build it the FDE way: thin, safe, end-to-end.
🏗️ Not into DevOps? Pick a different projectThe DevOps agent is one of six in-demand projects — support bots, code review, document intelligence, data analyst (text-to-SQL), and deep research are all in the Project Gallery, each a full design chapter mapped to these same chapters. Same skills, different domain — choose the one that fits your goals.
© 2026 studybydoing.in · AI Engineering: Zero to Production · All rights reserved. · About · Privacy Policy · Terms · Contact
Educational content, provided as-is and without warranty. Code samples are examples — review, test, and adapt them before using in production. See the Terms of Use & Disclaimer. Use at your own risk.
© studybydoing.in