I am genuinely curious about this question and want to understand people's experiences. I'm a principal engineer with decades of experience, I am currently working on a massive codebase that does RPA (data entry automation for old software). I can get my agents to build complex features from scratch, However, if I don't describe exactly how it needs to be built (low level specs), or if I don't know how the current codebase needs to be updated to support this change things quickly get out of control. So in other words, if I trust the agents to make any design decisions on their own, then a lot of nuance will be completely missed, leading to major bugs or issues with the SLA.

To summarize my problem, I am finding that the work required to build detailed enough specs, review and evaluate AI changes is far more time consuming (and cognitively demanding) than implementing the features myself.

Where am I going wrong?