Part 17: Cutting Edge & What's Next
Optional depth after the core curriculum — concepts that are shaping 2026–2027, not prerequisites for shipping today.
In one line: The rest of the guide teaches durable engineering; this chapter names the frontier ideas worth watching so you can skim announcements without chasing hype.
You do not need this chapter to build and ship AI products. Chapters 1–16 already cover what works in production today. Read this when you want a structured map of what is emerging — agent harness engineering, retrieval that plans its own searches, evals that grade the whole journey not just the final answer, cheaper inference tricks, and a dated snapshot of research themes. Treat it like a magazine appendix: interesting, useful for interviews and long-term bets, clearly perishable on the details.
How this chapter differs from Chapter 14
The frontier — and how to future-proof (Chapter 14) is a short pattern inside production patterns: durable skills vs. 2026 hype in one sitting. This chapter goes deeper on a handful of engineering shifts that are moving from research labs into production stacks — with links back to the foundations you already learned.
| Chapter 14 (Patterns) | Chapter 17 (This chapter) |
|---|---|
| One-page orientation | Five focused lessons + research radar |
| What to ignore vs. adopt | How emerging systems are built |
| Evergreen meta-skill | Dated snapshot on the research page |
What you'll learn
- Agent harness engineering — The orchestration layer around the model: memory, tool routing, budgets, and why the harness often matters more than the base model.
- Agentic RAG & memory — Retrieval that loops: when to search again, what to remember across turns, and how it differs from one-shot RAG.
- Trajectory & process evals — Judging agent steps, not just final answers — the eval shift agents force on you.
- Efficient models & test-time compute — Hybrid architectures, reasoning budgets, and the cost/latency tradeoffs that decide which tier runs where.
- Research radar (June 2026) — A dated snapshot of active themes and anchor papers — companion to the model snapshot, but for research direction.
Finish with an optional checkpoint — self-check only; it does not gate the capstone or glossary.
When to read this
- After you've worked through the core arc (or at least foundations + eval + one workflow chapter).
- Before a system-design interview where you want to sound current without reciting paper titles.
- When a headline claims agents or RAG are solved — this chapter tells you what is actually hard.
What stays durable here
Even when model names rot, these shapes tend to stick:
- Production agents need a harness (limits, memory, tool policy) — not just a smarter model.
- Grounded systems increasingly use multi-step retrieval, not a single vector search.
- Agent quality is measured on trajectories, not one final string.
- Cost wins come from routing (small vs. reasoning vs. frontier) and test-time compute budgets, not one model for everything.
→ Start with Agent harness engineering.