Coding Agents Are Broken
a day ago
- Coding agents are fundamentally flawed due to their harnesses, not just the models, which often generate low-quality code.
- The harnesses remove beneficial friction (e.g., slowing down for careful decision-making) while adding harmful friction like cognitive overload from verbose outputs.
- Agents tend to reinforce poor coding standards through feedback loops, as they match surrounding code quality, which degrades over time.
- Attempts to make agents act as engineers (e.g., giving extensive context or design upfront) fail because they cannot make tradeoff decisions or understand outcomes.
- The solution is not one-size-fits-all; users must customize agent configuration to their workflow, reducing cognitive overhead and integrating decision-making.
- Key questions to address: where agents cause needless cognitive load, how to maintain standards, how to align agents with personal workflow, and how to incorporate human judgment.