Wink Pings

In Large Projects, No One Understands All the Code—and the Same Goes for AI-Written Code

Sean Goedecke says that partial understanding is the optimal solution in large systems. Baoyu extends this judgment to AI Agents: do not expect it to read the entire repository; tests, review, and screenshot evidence are the baseline for delivery.

Sean Goedecke wrote in his personal blog that in large software projects, it is normal—and the best achievable state—for programmers to understand only the part of the code they are responsible for. Baoyu extended this point in a tweet: the same judgment applies to AI Agents. Many people say AI cannot write good programs because AI cannot fully understand the codebase. But programmers on large projects also only understand a small piece.

Sean Goedecke 推文截图

When newcomers take over a project with a complicated history, they can only understand part of the code in the short term. This does not prevent them from maintaining the project. In the past, automated tests provided the backstop. When test coverage is sufficient, it is not easy for a newcomer's changes to bring down the system; if they break something, the tests will not pass. Code Review is also a means of knowledge transfer and can prevent reckless changes.

The same goes for getting an Agent to do work. Expecting it to understand the entire project within one conversation context is unrealistic. It can only ever partially understand the project. Quality relies on verification as the backstop. Automated tests are essential, but you cannot rely on them excessively—some tests are problematic themselves. Code Review is also hard to guarantee now: the amount of code is too large for people to review it all, so most code can only be reviewed by AI, while people review the most important parts.

Baoyu mentioned that one method he has used often recently is having the Agent take screenshots. When developing a new feature, provide screenshots of the new feature; when fixing a bug, provide before-and-after screenshot comparisons. He also found another use: have the Agent simulate API response data itself and dynamically generate screenshots for different scenarios without needing server-side coordination. Some users guessed that this might use Playwright's page.route to intercept requests and return fake JSON, or that the frontend has MSW set up. Sean Goedecke asked in a reply why not just hit the local API directly, so the model does not have to guess. This detail shows that verification methods are still being refined through frontline use.

Agent 截图对比示例

Here is an easily overlooked point. Partial understanding is normal, but it does not mean you can make arbitrary changes. Max Kanat-Alexander said in a reply that engineers need enough judgment about their changes to explain what the real problem is and why the solution is clearly correct. They cannot use 'the AI said it is fine' as an excuse. Sean replied to him that the difference between 60% and 80% understanding is huge. Pascal put it more directly: full understanding usually means the system is smaller than you think. RISHi said that work is never about being omniscient; it is about knowing where the boundaries of your understanding are before you start.

So when AI does poorly on large projects, it is not necessarily because it understands too little. Often it is because boundaries and verification were never written down. One comment said that if changing one piece of code requires understanding a bunch of unrelated modules, then the architecture design and module boundaries must be wrong. Another comment said that AI often does poorly on large projects because no one ever wrote down the boundaries in the first place. This judgment is more useful than debating whether AI is smart.

Large systems rely on local changes, interface contracts, and deterministic acceptance gates. Do not make omniscience the goal. Leave the Agent checkable evidence: tests must run, screenshots must be reviewed, and critical paths must pass. Just hearing it say it is fixed is not enough. Original tweet link: Sean Goedecke, Baoyu's interpretation: link.

发布时间: 2026-11-21 19:09