Agentic Engineering — A Practitioner Playbook for Production Software with AI Agents

Both have an LLM write the code. What separates them is who decides the code is acceptable, and where correctness is pinned down — in a persons head, or in a file that can be re-run. With a comparison table and the same task done both ways.

Contents

1. 01 Vibe Coding Versus Agentic Engineering Both have an LLM write the code. What separates them is who decides the code is acceptable, and where correctness is pinned down — in a persons head, or in a file that can be re-run. With a comparison table and the same task done both ways.
2. 02 The Core — Context, Spec, Loop, Gate, Verify The five principles that let an agent work unattended, why an LLM must never be the judge of its own correctness, the shape of the whole loop, and what a single iteration is made of.
3. 03 Repository Structure for Agentic Work A standard layout that lets an agent opening a fresh session read the repository and get it right immediately, a map of the control files across Claude Code, Codex CLI and Cursor, and the context precedence order for when rules conflict.
4. 04 The Control Files — AGENTS.md, Skills, Hooks, Rules, Memory Working agent control files you can copy today: an AGENTS.md agents actually obey, skills loaded only when relevant, hooks that enforce rules mechanically, memory that stops the team repeating itself, all four Cursor rule types, and a script that keeps rules identical across tools.
5. 05 Loop, Gate, and Red / Green The gate is the heart of the system. Make it weak and the agent learns to turn it green without making the program correct. A complete verify.sh, a script that catches gate-cheating, the red-green order that must be enforced, and a loop driver with a ceiling and stall detection.
6. 06 A/B Tests, QA, and Measuring User Harm The testing layers in an agentic setting — what each catches and how far an agent can go — the two very different meanings of A/B, the five levels of harm, guardrail metrics wired to automatic rollback, and shadow mode.
7. 07 Zero to Production — The Nine Stages A real feature added to a running system, from spec to rollout, naming who does each stage, how to write a spec an agent can genuinely work from, and how review changes when a person did not write the code.
8. 08 MCP — When to Use It, How to Design It, What to Watch How to decide between an MCP server and an ordinary script, seven design principles for tools an agent uses well, a real server example, and the prompt-injection exposure that arrives with tool output — a genuinely usable attack, not a hypothetical one.
9. 09 A Web Search Stack for Agents — Tavily and the Free Alternatives An agent that cannot reach the web guesses library versions and writes code from expired memory. This chapter separates the four jobs people conflate, explains credit-based pricing, covers the free options — SearXNG, Jina Reader, Crawl4AI — and gives an architecture that tries free first and falls back to paid.
10. 10 Key, Token and Secret Security in Agentic Work One principle: the agent should never see a secret value. This chapter dismantles the common approach of keeping keys in a .md file behind agent permissions, separates registry from value, covers runtime injection, nine defence layers, a two-way secret hook, and an egress allowlist.
11. 11 Git for a Team Where Everyone Has an AI When code per person rises three to ten times but review capacity does not, the bottleneck moves from writing to reviewing and merging. Six core rules, a commit trailer recording provenance, a PR template that requires a blast radius, and CI in four layers.
12. 12 Appendix — Checklists and Anti-patterns A one-time project setup checklist, a checklist before every merge, the twelve anti-patterns that come up most with their remedies, and the whole playbook condensed into a single page of principles.

This series is All Rights Reserved — free to read, but copying, downloading, or republishing requires written permission.