(Im)Paired Programming: Coding Agents Improve Productivity but Harm Understanding
Source
Nishant Balepur, Connor Baumler, Valerie Chen, Eunsol Choi, Rachel Rudinger, Jordan Lee Boyd-Graber
Author
Nishant Balepur, Connor Baumler, Valerie Chen, Eunsol Choi, Rachel Rudinger, Jordan Lee Boyd-Graber
Date
Key takeaways · AI-distilled
54 students built a website with either a coding AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → that edits their code or a chatbot they had to copy from. The agent group finished better but scored worse on comprehension questions — and then could not extend their own code once the agent was taken away.
The damage tracks how you interact, not just whether you use an agent: copy-pasting prompts and auto-accepting edits are the specific behaviors linked to lower comprehension. Effort spent reading and steering is what preserves understanding.
Users kept preferring the agent even after reporting they understood their code less. Speed and ease win the preference vote, so you cannot rely on developers self-selecting away from the tool that is quietly costing them.
The suggested fixes are product-side, not willpower-side: discourage low-effort prompting, make agents emit readable code, and design interactions that force active engagement — because 'human in the loop' means nothing if the human stopped understanding.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
Why it matters
It puts evidence behind the suspicion that delegating edits to an agent quietly erodes the reviewer's grasp of their own codebase.
Key quotes
“Coding agents (e.g., Cursor) improve developer productivity by optimizing task completion, but shifting users from writing code to prompting and reviewing may harm their understanding, impeding oversight, learning, and communication.”
“While agents aid initial task completion, they harm users' code comprehension and thus do not prepare users to extend their code;”
“Despite self-reported weaker understanding, users still prefer coding agents because they are quick and easy to use.”
“While users stay in the loop for coding workflows, understanding should not be forgotten.”
videoStop babysitting your agents... — Brandon Waselnuk, UnblockedAI Engineer
articleBeyond code generation: rethinking engineering productivity in the age of AI agentsIlya Yakovlev,Andrew Cheung,Binoy Dash,Simran Jumani,Dmitriy Meyerzon,Mark Breitenbach,Ishan Mishra,Kazuaki Okumura