Optimize your developer environment

With an increased push to use AI for everyday work it might first seem that we can now easily scale developers’ productivity by 2x, 5x and beyond. And indeed, I see a significant boost of developers’ output all around me. However, I noticed that AI usage hides significant inefficiencies in our developer processes and most of the tokens we spend are not spent on the actual meaningful work. Instead, we spend tokens to overcome developer environment friction.

Indeed, tools like /loop and autonomous agents can spin for hours trying to:

figure out how to correctly run tests
query your CI system to parse logs of failed stages
try to fix flaky tests in unrelated PRs
use weird git commands to deal with stacked PRs

This all steals time, money, and opportunities to spend these tokens and compute on what actually matters. Unlike humans, agents don’t learn, so they are going to repeat these inefficient steps over and over again. I routinely see 70-90% of all the tokens consumed by devs spent inefficiently.

Don’t get me wrong, it’s still probably better and cheaper to throw AI and compute at the inefficient process instead of humans. GPUs don’t really care what they are working on and Anthropic will happily accept your payment. But unlike humans, AI is not going to complain. And the worst part is that once humans are removed from the process, no one is going to notice if things degrade. We’ll just wait longer and longer and pay more and more for tokens and compute. AI is not going to fix your slow builds while working on a features.

It’s in your best interest to monitor how much of human and AI time is spent on inefficient dev/babysit and try to get this number down. Sharpening your axe is still important even if the axe is flying around autonomously.

Practical steps:

Analyze your AI agent sessions and look for inefficient steps repeated over and over. You can even use agents for that!
Add metrics surrounding your developer environment commands. You should know P95s of your environment setup, dependency installation, typecheck, build and test times. Yes, even on local machines.
Give your agents enough context to follow efficient practices. Remember that agents don’t learn, so you need to explicitly provide context and tips.
If your codebase is large and complex, consider integrating a build system

If you are a manager, consider dedicating time and resources to improve your internal processes. Remove blockers to unlock these sweet productivity gains for your organization.