5 Skills I Still Learn by Hand While Agents Write Code

I measured the portion of my code that I still write myself versus that generated by agents. Surprisingly, about 80% of added lines originated from agent sessions. However, the commits that needed fixing, reverting, or debugging were primarily from that 80%. The remaining 20% I wrote by hand, such as interfaces, security checks, and test fixtures, were mostly accepted without modifications.

This experience led me to identify five essential skills that I still practice manually, even as agents handle more code. These skills are:

1. Quickly reading diffs: Reviewing code written by others in large volumes requires the ability to identify dangerous patterns, unchecked inputs, state mutations, and missing error handling paths.

2. Breaking work into small, testable units: Agents excel at handling small, well-specified tasks, but large, vague ones pose challenges. Learning to decompose work into manageable components enhances the success of agent-assisted development.

3. Debugging from first principles: When an agent's fix fails, or its suggestions do not resolve the issue, developers must analyze logs, perform bisecting, and form hypotheses. This skill was crucial in resolving a Postgres MCP bypass issue in a recent case.

4. Security fundamentals: Agents often reproduce vulnerable patterns present in their training data, potentially leading to security breaches. Understanding and writing insecure code helps developers recognize and prevent such issues during code reviews.

5. Learning from experience: Despite the growing influence of agents, developing intuition through personal experience remains invaluable. Building systems, experiencing architectural challenges, and understanding context management are essential skills that agents cannot replicate.