𝗧𝗵𝗲 𝗔𝗴𝗲𝗻𝘁 𝗗𝗶𝗱 𝗘𝘅𝗮𝗰𝘁𝗹𝘆 𝗪𝗵𝗮𝘁 𝗜 𝗔𝘀𝗸𝗲𝗱 𝗔𝗻𝗱 𝗧𝗵𝗮𝘁 𝗪𝗮𝘀 𝗧𝗵𝗲 𝗣𝗿𝗼𝗯𝗹𝗲𝗺

𝗧𝗵𝗲 𝗔𝗴𝗲𝗻𝘁 𝗗𝗶𝗱 𝗘𝘅𝗮𝗰𝘁𝗹𝘆 𝗪𝗵𝗮𝘁 𝗜 𝗔𝘀𝗸𝗲𝗱 𝗔𝗻𝗱 𝗧𝗵𝗮𝘁 𝗪𝗮𝘀 𝗧𝗵𝗲 𝗣𝗿𝗼𝗯𝗹𝗲𝗺

AI coding agents changed everything in 2024.

They read your codebase. They follow your naming patterns. They understand your domain.

I thought this meant a two-week task would take two days. I was wrong.

I gave an agent a large task: build a global event mechanism for a React Native app. I created a plan. I approved the plan.

The agent followed the plan perfectly. That was the problem.

The plan missed critical details. It did not account for components loaded conditionally. It did not map the impact on the whole system.

The agent solved the visible problems but left invisible ones behind.

Here is what happened:

• The UI broke in areas the plan ignored. • The agent added duplicate event handlers. • The agent invented new patterns using Context and Redux that I did not ask for.

The agent did not fail because of bad code. It failed because of a lack of constraints.

When you do not set a rule, the AI makes a guess. Its guess is often wrong.

I tried to fix it through chat. I added code on top of code. The chat thread became too long. I started new chats. The cycle repeated.

By day three, the code was 75% done but fragile. There was no record of what worked or what failed.

I realized the issue was not the tool. It was the role.

An AI can write code. It cannot decide what the work actually is.

Human review catches mistakes after they happen. But human judgment must define the work before it starts.

You do not need a better tool. You need a contract.

I built a seven-phase workflow. A human must approve every handoff before the next phase starts. This ensures the work stays within the defined boundaries.

Source: https://dev.to/jeelvankhede/the-agent-did-exactly-what-i-asked-and-that-was-the-problem-1hek

Optional learning community: https://t.me/GyaanSetuAi