Why Agentic Coding Workflows Can Undermine Engineer Output

Agentic coding—letting an LLM autonomously plan, write, and iterate on code with minimal human checkpoints—has attracted serious adoption. The argument for it is straightforward: reduce friction, compress cycle time, ship faster.

The counterargument, laid out in the Lars Faye piece, is that the tradeoff is less visible. When engineers offload not just boilerplate but the reasoning loops that produce boilerplate, they lose the feedback that builds system-level intuition. The code ships. The understanding does not.

This matters in maintenance. A codebase where agentic tooling made the decisions is a codebase the engineer cannot fully debug under pressure. The agent is not available at 2am when the service is down. The engineer is.

There is also a compounding effect. Junior engineers who skip the struggle phase of building something never develop the mental model required to catch what agents miss. Agents miss edge cases, security implications, and context that lives outside the prompt. That context is recoverable—but only if someone on the team holds it.

For senior engineers, the trap is subtler. Agentic tools feel productive because output volume rises. But output volume is not the same as delivery quality or system reliability. The signal degrades quietly.

The practical implication is not to avoid AI-assisted coding. It is to be deliberate about where the handoff happens. Using an LLM to generate a function body after you have written the signature and thought through the interface is different from asking it to design the interface. The first keeps the engineer in the reasoning loop. The second removes them from it.

Solo founders face a specific version of this risk. When you are the only person who understands the system, outsourcing that understanding to an agent is not acceleration—it is accumulating a debt with no one to collect it from but yourself.

The question worth sitting with: are you using agentic coding to move faster, or to avoid thinking?

Source