Method
Agentic engineering: AI under engineering control
AI speeds up execution. It replaces neither design, nor review, nor responsibility for the result.
Agentic engineering: AI under an engineer's control
Engineer first. AI executes; I design, decide, review and prove.
I do not delegate my understanding: every change is reviewed, and I keep the final say.
| Vibe coding | My practice |
|---|---|
| Accepting code without reading it | Versioned specs before any generation |
| Improvised prompts | A plan I approve |
| “It seems to work” | Systematic code review |
| Invisible debt | Tests first, tooled acceptance testing, evidence in production |
Six principles, each one verifiable
Scope before generating
Versioned specs split into deliverables with acceptance criteria; one instruction contract per repository.
Evidence: A 28-deliverable roadmap drove the SSO rollout across 9 applications.
Orchestrate
A plan approved before any code, specialized sub-agents, worktrees to parallelize without conflict.
Evidence: A fresh session executes another one's plan: no confirmation bias.
Set guardrails
Deterministic hooks, restricted permissions, no secret handed to AI, explicit human escalation.
Evidence: Guardrails are technical, not just instructions.
Prove it
Tests written first, then multi-agent testing in a real browser; every verdict is re-measured.
Evidence: 5 campaigns, over 800 cases, up to 11 agents in parallel.
Keep the final say
Every change reviewed, architecture and security calls made by me. No code I do not understand gets merged.
Evidence: I remain accountable for what goes to production.
Capitalize and stay ahead
Reusable skills, a wiki updated in the same patch, memory shared between agents.
Evidence: Active watch: spec-driven development, MCP, Claude Code, Codex.
From need to production
- 1
Business need & scoping
- 2
Versioned spec & roadmap
- 3
Approved plan
Human sign-off - 4
Implementation by agents (worktrees)
- 5
Code review & tests
Human sign-off - 6
Multi-agent acceptance testing (browser)
- 7
Fix loop
- 8
Sign-off & production release
Human sign-off - 9
Capitalization (skills, memory)
Tools
- Claude Code (Opus, Sonnet)
- OpenAI Codex
- Skills
- Hooks
- Subagents
- MCP (Chrome DevTools, Cloudflare, Resend, Vercel)
- Git worktrees
- AGENTS.md
- Hindsight
- GitHub Actions
Industrializing AI within a team
Shared rules rather than individual habits.
Shared conventions
An instruction contract per repository, versioned specs and a common definition of done.
Team skills
Reusable workflows (feature, acceptance testing, diagnostics) versioned with the code.
Guardrails
Hooks, permissions and mandatory reviews: a technical frame, not just a declared one.
Upskilling
Helping developers move from improvised prompting to a measurable engineering practice.
Frequent question
Why not simply use an AI or no-code tool?
Great for prototyping. Here is what changes when the software has to run every day.
A prototype is not a system
Generating a screen is easy. Connecting your software, handling permissions and edge cases, migrating your data is not.
Production has its own requirements
Authentication, backups, GDPR, monitoring, tests: what separates a demo from a tool you can rely on.
You stay the owner, and free to move
Standard, documented code hosted on your own accounts, which any developer can pick up.
A project to scope, or a team to equip?
Let's talk about your context: I will tell you plainly what AI changes, and what it does not.