The Road to Agent Engineer and Evangelist — A Long-Form Companion to the Learning Paths

Date: October 19, 2026 · Companion to the 🛠️ Agent Builder / 📢 Agent Evangelist paths

Executive Summary

This article is the long-form advice supplement for the two structured learning paths at Agent Academy. The on-site "key points" are the skeleton; this article answers what the skeleton cannot: Are these videos and courses actually worth your time? In what order? How deep is deep enough? And which traps does nearly every learner fall into?

All course content, instructors, and status facts cited here were verified against official course pages on 2026-10-19. Where public verification was not possible (e.g., the current upload cadence of individual creators), this is stated explicitly rather than asserted.

1. The Resource Map: Key Facts About the Main Courses and Videos

Learning resources expire, get revised, and close their certificate windows. Before you set off, confirm the map you are holding is still valid.

1.1 Structured courses (universities / official)

1.2 Video resources (by purpose)

1.3 Official documentation and must-read articles

• Anthropic · Building Effective Agents (Dec 19, 2024): draws the architectural line between workflows and agents, arguing that the most successful implementations use simple, composable patterns rather than complex frameworks. The most cited article on this site — re-read it at every stage.
• Model Context Protocol official docs: MCP is an open standard for connecting AI applications to external systems, described in its own docs as "like a USB-C port for AI applications" — build once, connect everywhere MCP is supported.
• OpenAI Cookbook: official examples for structured output (JSON mode) and function calling.
• Anthropic Prompt Engineering docs: widely considered the most systematic prompt-engineering methodology.
• LangGraph official docs: the de facto standard for state-machine agent orchestration.
• LangSmith / Langfuse, OpenAI Evals: the evaluation-and-tracing toolset.

1.4 Books

• AI Engineering (Chip Huyen, O'Reilly 2025): the chapters on agent architecture and evaluation deserve close reading.
• TED Talks: The Official TED Guide to Public Speaking (Chris Anderson): single through-line + concrete stories.
• The Sense of Style (Steven Pinker): classic style as the antidote to AI-flavored boilerplate.

2. Ten Pieces of Advice for Builders

1. Spend the first month on foundations — don't touch frameworks. Understand why the Transformer is an attention-weighted information filter, why the context window is the agent's working memory, and every parameter of messages / temperature / tools. The only exit test: hand-write a tool-calling conversational agent in under 100 lines, no framework. If you can't, you haven't earned the framework yet.

2. Token awareness separates engineers from players. Every prompt token is cost. Learn to estimate: system prompt size, what RAG injects, what one ReAct loop consumes, multiplied by daily active users. A build-from-scratch course like CS336 matters because it gives you an internal ledger of what the model is doing.

3. Tool design matters more than model choice. Most agent failures come from tool design: ambiguous parameters, mixed responsibilities, LLM-hostile return formats, no timeouts or retries. Write every tool as an SOP for a brilliant but forgetful new hire — the tool description is their entire onboarding.

4. Don't reach for a vector database on day one. Memory has three layers: short-term (context), long-term (vector store), task state (database). Before vectorizing anything, answer: what exactly will you retrieve, and where in the prompt will it land? No answer means an expensive, undebuggable landfill.

5. Re-read the Anthropic article at every stage. Building Effective Agents changes meaning as you progress: the workflow-vs-agent taxonomy at Stage 1, the "don't reach for frameworks" warning at Stage 3, the philosophy of simple composability at Stage 5. One article read five times beats five articles read once.

6. Frameworks are leverage — and liability. LangGraph models agents as state machines; CrewAI fits role-based division of labor. Frameworks solve orchestration while adding black boxes and debugging cost. The test: can you draw the full execution graph your framework generates for you? If not, go back to hand-writing.

7. Multi-agent is not "more people, more power." Every new agent adds an error-propagation surface. Push a single agent to its limits first; split only when the bottleneck is provably "one context can't hold this" or "one role can't switch fast enough." Human-in-the-loop at critical nodes is the line between a toy and production.

8. MCP is currently the highest-ROI skill on the path. The USB-C analogy is not marketing: wrap an internal API as an MCP server once, and every MCP-capable client — Claude, VS Code, Cursor — can use it. Security red lines: least privilege, read-only by default, second confirmation for sensitive operations.

9. No evaluation, no iteration. Build a golden set of 20 cases first and run it on every change — that one habit is worth more than ten blog posts. LLM-as-judge has biases; hand-check critical scenarios; record full traces. The cost triad: model tiering, caching, context trimming.

10. Milestones are not optional. Skipping milestones is the most common failure mode on this path — debt from the first three stages comes due, with interest, at stages four and five.

3. Six Pieces of Advice for Evangelists

1. Your biggest risk is being wrong, not being boring. Exit test for fundamentals: explain the Transformer to three audiences — a child (analogy), a product manager (value), an engineer (mechanism). Missing any one means it isn't time yet.

2. Keep a fact ledger. Every claim traceable to a paper or official doc, with a confidence note. Follow primary sources. Credibility comes from not overselling — understanding cost and limits matters more than understanding capability.

3. Structure beats eloquence. The golden structure: problem → wrong approaches → new idea → demo → call to action. Leave exactly one takeaway per talk. Live-demo iron rule: always have a recording as backup; only run on stage what has succeeded three times before.

4. Study the masters; don't imitate them. Karpathy's pacing, 3Blue1Brown's animated intuition, Mu Li's paper structure, David Malan's stage control — dissect these, don't copy them. Your style grows out of your own hands-on projects.

5. Frequency beats virality. One piece per week for twelve weeks. The flywheel: real project → tutorial → Q&A → collected questions → next topic. Turn every answer into a FAQ entry — your best topic pool.

6. Enterprise evangelism is translation. Translate agent capability into the client's language of cost and efficiency. Prepare three demos: code for engineers, ROI for management, experience for end users. Record every case as background → approach → outcome → failure modes. Failure modes earn the most trust. Never promise what you can't deliver; never relay what you haven't verified.

4. Four Pieces of Advice Common to Both Paths

1. Output drives input. Milestone projects and published essays are the same thing: verifiable output that closes the learning loop.
2. Certificate windows close; capability doesn't. Berkeley MOOC certificates have all been issued, but the videos remain; the HF course still offers free certification as of this writing. Start early if you want the paper — never make the paper the goal.
3. Resources expire; keep the map current. Top courses get new editions every year (CS336 is at Spring 2026; 6.S191 is rewritten annually), and framework docs change monthly. Re-validate your learning list quarterly.
4. Loneliness is the biggest enemy. Find a community where you can ask questions and be corrected. The self-learner's greatest risk is learning it wrong with nobody telling you.

5. Quick Reference: Common Misconceptions

6. A 90-Day Action Plan (Builder Track Example)

Evangelists can compress accordingly: weeks 1–4 fundamentals + a 3,000-word essay; weeks 5–8 methodology + a 10-minute recorded video; weeks 9–12 four consecutive weekly publications; then talks and enterprise evangelism.

Closing

There are no secrets on this path: principle courses, official documentation, deliberate practice, verifiable output. The only variable is whether you are still moving forward week by week on day 90 and day 180.

The route is drawn, the resources are verified, the advice is delivered. The rest is beyond what any website can do for you.

Chapters: 9 Milestone Companions

This article is the overview; each milestone has its own targeted long-form companion — read the one for the stage you are on:

• 🛠️ s1 Foundations: Your First Agent Is a Loop, Not a Framework
• 🛠️ s2 Core Agent Paradigms: Make Retrieval Prove Itself
• 🛠️ s3 Frameworks and Orchestration: Earn the Framework by Drawing the Graph
• 🛠️ s4 MCP and System Integration: One Server, Every Client
• 🛠️ s5 Reliability and Production: No Eval, No Iteration
• 📢 s1 Foundations: Be Right Before You Are Heard
• 📢 s2 Method: Structure Beats Eloquence
• 📢 s3 Platform: The Flywheel Beats the Viral Hit
• 📢 s4 Influence: Translation Is the Job

Verification notes: factual claims about course status, instructors, and protocol descriptions were verified on 2026-10-19 against llmagents-learning.org, stanford-cs336.github.io, web.stanford.edu/class/cs224n, introtodeeplearning.com, huggingface.co/learn/agents-course, github.com/microsoft/ai-agents-for-beginners, anthropic.com/research/building-effective-agents, modelcontextprotocol.io. Reputation descriptions of video creators reflect long-standing public standing and were not individually re-verified; readers should judge their timeliness independently.