Traditional Coding vs Agentic Coding: The Flow State Problem

Manually coding offers a tight feedback loop that keeps you engaged in the process. You think, write, run, observe, and adjust, moving through a cycle of decisions that builds familiarity with the problem at hand. With agentic coding, this loop is disrupted. You provide an idea and then wait for the agent to work, checking your email or moving to another project.

The agent produces code, but you don't see the entire process unfold. The output is exciting, but it can also scatter your attention. Instead of focusing on one problem for an extended period, you repeatedly switch between different tasks. This fragmented attention makes it difficult to maintain a deep understanding of any single aspect of the project.

The agent's work can introduce uncertainty, as you have to remember the details of each previous task to evaluate the agent's output. Notifications can further break your focus, causing you to switch tasks based on alerts rather than reaching a natural stopping point. The loss of ownership when you hand off a large task to an agent can be disorienting.

You contributed the idea and planned the work, but the agent handles many decisions. When changes are needed, you may not fully grasp the underlying assumptions, making it harder to adjust the codebase effectively. A more involved workflow that keeps you connected to the code involves breaking tasks into smaller, manageable parts.

Instead of requesting a complete authentication system, start with the user model and test it before moving on to the next piece. This approach allows you to understand and review each change before proceeding. By keeping tasks small and focused, you maintain a strong connection to the project and experience the satisfaction of solving problems one step at a time.