Agentic Code Generation and Developer Fatigue
Apply friction to agentic sessions: write initial code, then review, ask questions, etc.
Implement friction steps in your development workflow to mitigate agentic fatigue.
Summary
Agentic code generation can leave developers feeling mentally exhausted, as the tool often bypasses the internal processes that reinforce skill retention.
The author explains how human memory works in three layers—short‑term, working, and long‑term—and how reading code engages all of them simultaneously. Default LLM usage feels like a slot machine: you pull a lever and get a solution, but you lose the learning loop. To counter this, the author suggests adding friction back into the workflow: write the initial implementation yourself, then let the agent review it comment‑by‑comment. Use the agent to ask clarifying questions and pull relevant documentation, or to propose multiple approaches for you to critique. Discuss the agent’s output with a colleague, and only invoke the tool after spending at least 20 minutes on the problem. The author also recommends reading academic papers and reimplementing fundamental data structures to solidify foundations. While these practices slow down short‑term speed, they ultimately improve long‑term proficiency with LLMs.
Key changes
- Write the initial implementation yourself before asking the agent
- Review the agent’s code comment‑by‑comment
- Use the agent to ask clarifying questions and pull relevant documentation
- Ask the agent to propose two approaches and critique the alternatives
- Discuss the agent’s implementation with another person
- Only invoke the agent after spending at least 20 minutes on the problem
- Read academic papers and reimplement fundamental data structures
- Add friction to mitigate agentic fatigue and reinforce learning