Agentic Engineering
Agentic Engineering & Spec-Driven Development
Agentic engineering is how software teams ship with AI agents without losing control of quality. Instead of prompting until something works, the team writes a spec, the agent turns it into a plan and small tasks, and humans approve each step before code is merged. I help engineering teams set this up on the tools they already use, like GitHub Copilot in VS Code agent mode and Claude Code.
Agentic Engineering
Vibe coding vs agentic engineering
Vibe coding is great for prototypes: describe what you want, accept what the model writes, repeat. It breaks down in a real codebase, where nobody can say why the code looks the way it does, reviews turn into guesswork and every new session starts from zero. Agentic engineering keeps the speed of agents and adds the structure a team needs: written intent, small steps, clear gates and checks that run every time.
| Vibe coding | Agentic engineering | |
|---|---|---|
| Starting point | A prompt | A written spec |
| Context | Whatever is in the chat | Specs and agent guidance files in the repo |
| Unit of work | One big change | Small tasks from a reviewed plan |
| Human role | Accept or retry | Approve the spec, the plan and each merge |
| Quality check | It runs | Tests, lint, build and review on every task |
| Memory | Lost between sessions | Kept in the repo for the next agent and the next teammate |
Agentic Engineering
Spec > plan > tasks > code
- 01
Spec
What we are building and why, the constraints, and how we will know it works. Owned by a human, often drafted with the agent.
- 02
Plan
The technical approach, the files involved and the risks. The agent proposes it; a human approves it.
- 03
Tasks
The plan broken into small, ordered, testable steps that an agent can finish and a reviewer can read.
- 04
Code
The agent implements one task at a time, runs the checks and hands back a diff to review.
Agentic Engineering
How I run agentic engineering in a team
- 01
Specs in the repo
Every feature starts as a short spec checked in next to the code, so intent gets reviewed before any code exists.
- 02
Agent guidance files
AGENTS.md, Copilot instructions and Claude Code skills tell every agent how the codebase works: conventions, commands, boundaries and what not to touch.
- 03
Human gates
Agents stop for approval at the points that matter: the spec, the plan, and before anything is merged or deployed.
- 04
Review flow
Small diffs, one task each, reviewed like any teammate's pull request, with the spec as the reference.
- 05
Verification
Tests, lint, type checks and a build run on every task, and UI changes get checked in a real browser before anyone calls them done.
Agentic Engineering
I built these guardrails before AI
At MyHeritage I built the developer tooling and CI/CD pipelines that let engineers ship code securely to millions of users, and made sure nothing broke along the way. That was years before AI coding agents. Agentic engineering is the same job with a new kind of developer: the gates, checks and review flow that kept human code safe now keep agent code safe.
Agentic Engineering
Tool-agnostic: GitHub Copilot, Claude Code, Spec Kit
GitHub Copilot (VS Code agent mode)
Repository instructions, prompt files and custom agents give Copilot the same context every time. Agent mode in VS Code works through the plan task by task, and the Copilot coding agent can pick up well-specified issues.
Claude Code
CLAUDE.md and AGENTS.md hold the project rules, skills package repeatable workflows, and subagents and hooks enforce the gates and checks.
Spec Kit
GitHub's open-source Spec Kit gives the flow a ready-made shape: a project constitution, then specify, plan, tasks and implement commands that work across Copilot, Claude Code and other agents.
The method matters more than the tool. The same specs and guidance files carry across agents.
Agentic Engineering
AI-DLC and how it maps to spec-driven development
AI-DLC, the AI-Driven Development Lifecycle, is a methodology from AWS for building software with AI at the center of the process and humans in charge of the decisions. Work moves through three phases: Inception (requirements, planning and architecture), Construction (design, code generation, build and test) and Operations (deployment and monitoring). At each step the AI asks clarifying questions, proposes a plan and waits for a human to approve it, and every decision is logged.
AWS publishes open-source AI-DLC workflows that run inside coding agents such as GitHub Copilot, Claude Code, Amazon Q Developer and Kiro. The lifecycle lines up closely with the spec-driven approach on this page:
| AI-DLC | Spec-driven development |
|---|---|
| Inception | Spec and plan, approved by a human |
| Construction | Tasks, code and review, one unit at a time |
| Operations | Verification, deploy and monitoring |
| Approval at every phase | Human gates on the spec, the plan and each merge |
| Audit trail of decisions | Specs, plans and guidance files kept in the repo |
The two fit together: AI-DLC gives the lifecycle and the approval points, and spec-driven development gives the day-to-day workflow inside each phase.
AI-DLC overviewAgentic Engineering
FAQ
What is agentic engineering?
Agentic engineering is a way of building software where AI coding agents do much of the implementation inside a structured process: work starts from a written spec, is broken into a reviewed plan and small tasks, and passes human gates and automated checks before it ships. It is the disciplined, team-ready alternative to vibe coding.
What is spec-driven development?
Spec-driven development means the spec, not the prompt, is the source of truth. A short written spec describes what to build, why, and how to verify it. The agent turns it into a plan, then tasks, then code, and each step is reviewed against the spec.
What is the difference between spec-driven development and vibe coding?
Vibe coding is prompt, accept, repeat. It is fast for prototypes but hard to review, repeat or maintain. Spec-driven development keeps the agent's speed but adds written intent, small reviewable steps and automated checks, so the result is code a team can own.
What is AI-DLC?
AI-DLC (AI-Driven Development Lifecycle) is AWS's methodology for AI-native software delivery. Work runs through three phases, Inception, Construction and Operations, where AI proposes plans and does the work while humans approve each step. It maps directly onto spec, plan, tasks and code.
Does this work with GitHub Copilot?
Yes. GitHub Copilot in VS Code agent mode supports repository instructions, prompt files and custom agents, which is everything spec-driven development needs. GitHub's Spec Kit and AWS's AI-DLC workflows both support Copilot.
Does it work with Claude Code and other agents?
Yes. The approach is tool-agnostic. Claude Code uses CLAUDE.md, AGENTS.md, skills and hooks for the same purpose, and the same specs and guidance files work across agents.
What does an Agentic Engineering Enablement engagement include?
Setting up the workflow in your repos (spec templates, agent guidance files, human gates and the review flow), wiring in the checks, and training your team to run it on their own.
Bring agentic engineering to your team
Book a 15 minute call. We'll look at how your team ships today and what a rollout of spec-driven development would look like: setup, guidance files, gates, review flow and team training.