Curious to see what you do
- Let the agents cook, and not see the code19%, 5 votes
- Let the agents cook, but still review67%, 18 votes
- Code by hand15%, 4 votes
27 votesFinal results
Commit
993ff1c
My biggest issue today is that after an agent cooks, the review agent gives feedback to tune it. I often end up with a huge, over-engineered mess. So have the cut out the fat.
Runar also asked this, but can you expand on your setup. I just usually use one model, and then review it myself.
I just open a new tab in OpenCode with a brand-new agent. If the work involves a lot of JavaScript, I give it a JavaScript skill and then just tell it to "review the uncommitted work in this project".
Feed that back to the main agent with the instruction:
These bugs may be correct/incorrect the solutions may be good/bad. Inestigate, research and develop a plan to fix the relevant ones.commit 993ff1c