Hello, this is Pragmatic Developer.
Today, I would like to share a story from my practical experience. I have organized the thoughts I've been having lately while reviewing PRs created by AI.
When you assign a task to an AI agent, it finds the relevant files, implements the solution, and even attaches test codes, resulting in quite a large set of changes. However, the speed at which a human understands code remains almost the same. Within an eight-hour workday, we have to attend meetings, discuss planning, and do our own development, yet PRs keep piling up five or six at a time.
That is why I look at the thought process behind the work before looking at the diff—what problem they were trying to solve, why they chose this implementation method, and what they verified and how.
The video covers the following topics.
- Why code reviews become a bottleneck in the era of AI coding
- The logic of the work and the basis for verification, which are checked before the code itself
- 5 things that must be in a good PR (Problem / Approach / Scope / Verification / Remaining Risks)
- Criteria for allocating review focus based on risk level
- The change in the reviewer's role, from someone who finds bugs to someone who verifies the grounds for validation
- 4 things that new and junior developers should train on right now
Video Link
Points that connect with the lecture
In the end, this video is about "how to control the outputs created by AI."
Requiring evidence during the review stage is one approach, but it is more fundamental to design the work environment so that the agent produces verifiable results from the start. The feedback layer among the four layers of harness engineering covered in the lecture addresses this exact issue.
This is the part where you design which validations to run automatically, how to feed failures back to the agent, and what to leave for human verification. It also includes a hands-on exercise on enforcing pre-commit validation using Hooks.
The mention in the video about "forcing validation details to be written in the PR body" is ultimately about setting up a human-side harness. Moving that same principle into the tools and control flow will improve the quality of the agent's work itself.
Please enjoy the video, and I will be back with more practical news.
Additional News
Lately, I have been thinking from various angles about how I can provide information that will truly be helpful to all of you. Although I haven't announced it publicly yet, I am preparing content that will be genuinely beneficial for current developers, aspiring developers, and job seekers (of course, it will be free for my students), so please stay tuned for more news 🙏
So, keep up the great work with your studies today!!