Prasanta Kumar Behera
AI-Assisted Infrastructure Engineering›01 · The AI-assisted ops landscape
Learning Hub / AI-Assisted Engineering / AI-Assisted Infrastructure Engineering

Lesson 01 of 14 · Modules

The AI-assisted ops landscape

A realistic map of AI in infrastructure work: chat assistants, agentic coding tools, tool integrations through MCP, and AIOps. Where they save real time today, how they fail (confident errors, stale knowledge, prompt injection), and the principle that runs through this track: AI proposes, tools and humans verify.

Beginner → Practitioner
Key wordsLLMsAI assistantsagentic toolsClaude CodeMCPAIOpshallucinationsverificationwhere AI helpswhere AI should not be trusted

The landscape

Category What it does well Typical risks
Chat assistants Explaining concepts and errors, drafting scripts/docs, converting formats Outdated or invented details
Agentic coding tools (e.g. Claude Code) Working across a repo: reading, editing, running tests, iterating Wrong changes at speed; over-broad permissions
Tool integrations (Model Context Protocol servers) Letting models query systems (clusters, tickets, metrics) through defined tools Excessive access; prompt injection via data
AIOps Anomaly hints, alert correlation, summarising incidents False positives, opaque reasoning

An AI assistant is like a very fast, very well-read new colleague. It has read most of the manuals, types instantly, and never gets tired, but it sometimes remembers things wrong and says them confidently. You'd happily let it draft things and suggest ideas; you'd check its work before it touches anything important.

Where it saves real hours today

  • Drafting: scripts, runbooks, RCA write-ups, documentation, commit messages, PR descriptions.
  • Translating: natural language ↔ jq, PromQL, LogQL, regex, awk; one config format to another.
  • Explaining: unfamiliar code, error messages, long logs, a Helm chart's values.
  • Scaffolding: tests, CI pipelines, Terraform modules, which you then review and verify.
  • Reviewing: a second pair of eyes on PRs for common risks (lesson 06).

How it fails

  • Confident errors (often called hallucinations): invented flags, wrong API versions, plausible but non-existent options.
  • Stale knowledge: models have training cut-offs; tools change fast.
  • Missing context: it doesn't know your environment unless you give it (and giving it has data-handling implications).
  • Prompt injection: text in logs, tickets or web pages can contain instructions that an agent might follow if it's not designed carefully (lesson 14).
  • Over-trust: the output looks right, so people skip review.

The principle of this track

AI proposes; deterministic tools and humans verify.

idea / question ─► AI draft (code, plan, diagnosis)
                       │
                       ▼
        verify: tests, linters, policy (OPA/Kyverno), dry runs, plans, docs, a human reviewer
                       │
                       ▼
               apply with normal change control

The rest of the track applies this to daily work (02), engineering practice (03), Claude Code automation (04), MCP (05), PR review (06), testing (07), IaC (08), diagnostics (09–11), agents (12), GPUs for AI workloads (13) and governance (14).

Try it: calibrate your trust

  1. Ask an assistant five questions about tools you know very well (e.g. specific kubectl flags, Terraform behaviours); mark each answer right, partly right, or wrong.
  2. Ask it to write a jq filter and a PromQL query for real data; test both.
  3. Ask about a feature released in the last few months; compare with the official changelog.
  4. Note which kinds of questions it got right reliably and which needed verification.
  5. Write your personal "trust but verify" checklist (three to five items).

Going deeper: adopting AI in a team

  • Agree on approved tools and data rules (what can be shared with which provider) before usage spreads informally.
  • Measure outcomes (time saved, defects, review load), not just usage.
  • Share prompts, skills and patterns that work, like any other engineering knowledge.

Recap

  • Four categories: chat assistants, agentic tools, tool integrations (MCP), AIOps.
  • Big wins in drafting, translating, explaining, scaffolding, reviewing.
  • Failure modes: confident errors, stale knowledge, missing context, prompt injection, over-trust.
  • AI proposes; tools and humans verify.

This site is a public version of my personal engineering knowledge hub. It intentionally excludes confidential company information and internal operational details.