Production engineering

25 hours · the parts that appear in job descriptions and in no other roadmap

Everything here answers one question: does this behave when it's real? Anyone can get a model to produce the right answer once. This pillar is about the difference between that and a system you'd put in front of customers — which is the thing hiring loops are actually trying to detect.

It's a small pillar with a high hit rate, because most of it is reading and applying judgement you already have rather than learning new machinery. It's also the pillar where an applied practitioner can sound obviously senior in about ten minutes of conversation.

Modules

1 · Agents, honestly · ~7h

Likely the strongest existing area for anyone arriving from applied LLM work — so the goal is sharpening and packaging, not learning. Re-reading framework tutorials is exactly the procrastination this roadmap warns about.

• Building Effective AI Agents — Anthropic · skim. The clearest statement of the workflow-vs-agent distinction and when each pattern applies. Primary source, and the vocabulary interviewers increasingly use.
• Model Context Protocol · work through. Anthropic released MCP in Nov 2024 and donated it to the Linux Foundation in Dec 2025; it's now adopted by OpenAI, Google and Microsoft and is effectively baseline knowledge for agentic work. Build one MCP server. An afternoon, and not knowing this in 2026 reads badly. Hugging Face's free MCP course is the structured version.

What to actually practise: arguing against an agent. The most convincing interview answer is usually "this should be three chained calls with a retry, not an agent" — with the cost, latency and debuggability reasoning behind it.

Done when: you have built and deployed one MCP server, and you can defend a single-agent, multi-agent, and plain-workflow design for the same problem, saying which you'd ship and why.

2 · Structured outputs and reliability · ~4h

The gap between a demo and a product is usually here: LLMs produce inconsistent JSON, malformed syntax and unpredictable field names, and it gets worse when a model version changes underneath you.

• Instructor · work through. Pydantic-based structured extraction with validation and retries built in; the de facto standard alongside Outlines for API-based and self-hosted models respectively.

The pattern worth internalising: enforce schema at generation time, validate semantics at the application boundary, and handle structural failures differently from semantic ones. A retry fixes malformed JSON; it does not fix a confidently wrong value.

Done when: you can explain why schema enforcement isn't validation, and what your system does when the model returns well-formed nonsense.

3 · Security and prompt injection · ~7h

The gap most likely to be exposed in an interview about agents, and the one this roadmap originally missed entirely.

• OWASP Top 10 for LLM Applications · work through. Prompt injection has held the top spot for two editions running. Read the whole list — excessive agency, system prompt leakage, improper output handling and unbounded consumption all describe failure modes you can hit this week.

The core insight, worth being able to state cleanly: LLMs process instructions and data on the same channel, so any untrusted content the model reads — a webpage, a document, a support ticket, a repo — is a potential instruction. There is no prompt that fixes this. What works is defence in depth: least privilege on tools and credentials, allowlisted actions, segregating untrusted content, output filtering, human approval for anything consequential, and audit logs.

Retrieval and agents make this sharply worse, which is precisely why it belongs next to those pillars rather than in a security appendix.

Done when: you can walk through how an indirect prompt injection reaches an agent that browses the web, and name four independent controls that limit the blast radius — without claiming any single one solves it.

4 · Cost engineering · ~4h

Named explicitly in 2026 job descriptions, and a fast way to sound like someone who has run something real. Teams applying the full stack of techniques report 60–80% reductions in token spend without quality loss.

The levers, roughly in order of return: model routing (cheapest model that clears the quality bar, per use case) · prompt caching (large cuts on repeated context) · semantic caching (serving near-duplicate queries) · batch tiers for anything not latency-bound · prompt compression · context budgets per request.

Do the arithmetic once for a system you know: tokens per request, requests per day, cost per month, and what each lever would save. Numbers you produced yourself are the ones you'll recall under pressure.

Done when: you can take an unfamiliar LLM feature and produce a monthly cost estimate plus the three changes you'd make first, with rough percentages.

5 · Observability · ~3h

The unglamorous half, and the half that shows up in senior loops.

Tracing spans across a multi-step chain · what to log when output is nondeterministic · latency budgets per stage · PII in prompts, logs and traces · sampling live traffic for continuous scoring · detecting regression before a customer reports it.

Done when: you can describe how you'd debug a quality regression reported by one customer, from trace to root cause, without hand-waving.

What I'm skipping

• Vendor tool tours. This landscape churns constantly and the concepts don't. Learn one tool well; know the category exists.
• Red-teaming as a specialism. Know the attack classes and the controls. Adversarial ML research is a different job.
• Framework comparison essays. Know the tradeoffs, skip the reading.