Modern Engineering Values
A few things that I want to highlight:
I am ineffective at running multiple agent sessions in the same project (with worktrees, copies, or other solutions).
Have the similar feeling, worktrees becoming harder to manage and eventually there is a strong desire to point agent in one worktree to another. At this point I am usually mentally done.
When fixing bugs, I point it at a repro or context and force it to first create a failing test. This significantly increases the likelihood of the model fixing the right problem in the right way. Otherwise models tend to fix the wrong thing, make the wrong fix, or write the wrong test.
Great advice and behavior I seen most competent models picking up automatically sometimes.
The best next step is to push to main. Unfortunately, that’s not how most teams operate, so once you enter the PR or collaboration phase you kind of lose most of the velocity that LLMs give you. I believe that all technical organizations will have to drastically change and break down barriers to getting things done if they want to be successful going forward.
Very strong agree. The answer how to effectively review code produced by another human+LLM remains to be elusive. Be much harsher? Makes no real sense and dick move. Rubberstamp? Cringe and performative theater. Look for exclusively the most serious issues like security etc?
Someone who deeply understands a product can now execute faster, while someone without context can generate much more noise. As a result, coordination becomes more expensive. I believe the most effective teams will be small, often two to three people, with clear ownership boundaries and isolated repositories.
Pizza-sized teams getting serious backing. It’s possible that boundary of more people added to the project resulting in mess becomes much lower. Doubly so in case of AI euphoric team that can go off rail in a matter of days
Strict guardrails and fast feedback loops are the difference between an agent completing their work in 1 minute or 60
It’s not obvious to me yet which constraints are really making some sense. Clippy-like universally acknowledged advice might be the best we have in the industry as of now but it’s not THAT much.
With agents, there is no longer a good reason to outsource core parts of your product to third-party dependencies if you can realistically own them yourself.
Not completely sure because one of reason we benefit from 3rd party dependencies is unreasonable (for us) amount of care and work being put to seemingly obvious or “straightforward” idea. Vendoring though might be adopted much more universally
Great observations and lots food for thought though.