How to Master AI-Assisted Coding with Agentic Engineering

Introduction

In the rapidly evolving world of software development, using AI to write code has moved beyond novelty into a core practice. But as Chris Parsons highlights in his updated guide—the third iteration of his approach—the key is not just asking an AI to generate code, but building a disciplined workflow around verification, automated checks, and continuous training. This how-to guide distills the essential steps for adopting agentic engineering, a method that treats AI as a collaborator that you train to produce reliable, ship-ready code. By following these steps, you'll shift your focus from "how fast can we build" to "how fast can we tell whether this is right." You'll learn to set up your environment, implement verification pipelines, and scale your expertise by teaching others. Let's dive in.

How to Master AI-Assisted Coding with Agentic Engineering — Source: martinfowler.com

What You Need

An AI coding tool – Chris recommends either Claude Code or Codex CLI. These provide a strong inner harness for agentic workflows.
A test framework – such as Jest, PyTest, or similar, to create automated tests for every change.
A type checker – like TypeScript, MyPy, or Flow, to catch structural errors early.
A continuous integration/continuous deployment (CI/CD) system – to run tests and checks automatically on each commit.
A code review platform – for human oversight where automated gates aren't sufficient.
Documentation habits – a documented approach for your AI prompts and feedback loops.
Time for training – iterative sessions where you teach the AI your coding standards.

Step-by-Step Guide

Step 1: Choose and Set Up Your AI Coding Tool

Select a tool that supports agentic workflows. Both Claude Code and Codex CLI offer robust inner harnesses—the scaffolding that allows the AI to call tools, run tests, and self-correct. Install the tool and configure it to have access to your project's repository. Ensure you understand the tool's capabilities for executing commands and returning results. The inner harness is a critical advantage because it lets the AI verify changes without human intervention.

Step 2: Build Verification Pipelines First

Before writing any AI-generated code, set up your verification infrastructure. Create automated tests that cover the key behaviors of your application. Set up a type checker to run on every file. Integrate these into your CI/CD system so that every pull request is automatically validated. The goal is to make verification as fast and frictionless as possible. As Chris says, "Build better review surfaces, not better prompts."

Step 3: Keep Changes Small and Documented

When you ask the AI to generate code, limit each request to a small, focused change. Large prompts lead to large diffs that are harder to verify. Instead, break your feature into small increments. Document each change thoroughly—include comments in the code and update any relevant design documents. This documentation trains the AI to understand your project's conventions over time.

Step 4: Implement Guardrails for Automated Checks

Set up automated gates that the AI's output must pass before it reaches your eyes. For example, configure the AI tool to run the test suite after each code generation, and to fix any failures automatically. Similarly, require that the code passes type checking and linting. These guardrails perform the bulk of verification, leaving only the most nuanced decisions to you. This is the shift from "verified by you" to "checked by tests, type checkers, and automated gates."

Step 5: Train the AI Through Iterative Feedback

As you review the AI's output, give specific feedback. If the code uses a wrong pattern, correct it and explain why. Over time, the AI will learn your preferences. Use a consistent set of prompts and feedback templates. This training accelerates with each iteration, so the diffs become right more often. As a senior engineer, your job evolves from approving diffs to shaping the AI's behavior. The way to stay valuable is to "train the AI so the diffs are right the first time."

Step 6: Shift Your Focus to Rapid Verification

Instead of measuring how many features you generate, measure how quickly you can verify whether a change is correct. A team that can generate five approaches and verify all five in an afternoon will outpace one that generates one approach and waits a week for feedback. Invest in making feedback loops fast: reduce test run times, parallelize checks, and use the AI's ability to self-verify against a realistic environment. The game is now speed of verification, not speed of generation.

Step 7: Scale by Teaching Others

The most important skill for an agentic engineer is the ability to pass your knowledge on to colleagues. Document your prompts, your harness configuration, and your feedback techniques. Create a training session or a shared artifact that teaches others how to use the same verification-first approach. This compounds your impact far beyond individual contributions. As Chris notes, "If you are a senior engineer worried that your job is quietly turning into approving diffs, the way out is to make yourself the person who shapes the harness."

Tips for Success

Prioritize verification speed over generation speed. The quickest loop wins.
Invest in the harness—your tool's inner scaffolding is where you'll get the biggest returns.
Make feedback instant wherever possible. Use the AI to run checks before asking a human.
Document everything, especially the rules you want the AI to follow. It pays off in future iterations.
Start with small, safe changes when training a new AI model or tool. Build trust gradually.
Automate the boring verification—leave only the high-judgment decisions for human review.
Embrace the shift: your role is now a coach and harness shaper, not just a code reviewer.

By following these steps, you'll harness the full potential of AI-assisted coding without sacrificing quality. The future is not about faster building—it's about faster knowing. Use these techniques to stay ahead.

Tags: