Taste in the Loop

Engineers are getting buried in AI-generated PRs full of slop. The cheapest way to prevent this is to give agents a good plan, before they write code. At ContextBridge, we use an open-source tool we built called PlanBridge. It makes iterating with your agent’s plan a lot nicer.

Every line of code incurs a maintenance cost (even if it’s small). Thus if AI lets us write code faster, we need to reduce maintenance costs by a proportional amount to preserve our productivity gains. This is hard to do in practice (as James points out). But, we’re starting to see nascent attempts at what we like to call “janitorial agents”; these are async background agents that automatically refactor, fix bugs and update docs (e.g. based on a cron schedule).

Shiv’s post will resonate with many engineers who have worked with AI long enough. With several memorable passages such as:

AI writes features, not architecture. The longer you let it drive without constraints, the worse the wreckage gets.

As AI models continue to improve, we believe spending time on good architecture and thoughtful planning will remain an important lever to steer agents towards better outcomes. Many seem to think that a linter and “tests” (which the agent wrote) are sufficient for keeping an agent on-track. They are not. Good engineering was never just about the code.