GitHub’s Post

Claude finally learned the ancient engineering principle: maybe don’t open twelve tabs.

The model cadence is wild. Positioning Sonnet 5.5 at well-scoped everyday work is smart, features and bug fixes are where devs spend most of their week, so that is where the compounding shows up fastest.

Fewer steps on well-scoped work, nice. What does it do when the issue isn't well scoped: ask, or finish the wrong thing faster? 🤔

Fewer steps and tool calls are promising if the resulting patches remain easy to review. It would be useful to compare failure recovery too: how often does the agent notice a failing test, correct its own change, and leave a concise explanation of what happened?

Exactly. For me, fewer steps would make the biggest difference too. The less there is to review and double-check, the easier it is to trust the output and keep moving.

Fewer tokens per task cuts costs in high-volume build pipelines. For automation running thousands of times, that margin compounds. Speed in iteration cycles is where it counts.

On the underspecified-task question: I'd include deliberately ambiguous issues in the eval, with “ask a clarifying question” as a valid outcome. Score whether the acceptance criteria were met, not just steps or tokens. Otherwise a model can look more efficient by skipping the question that would prevent the wrong change.

Fewer steps is the number I'd look at first. I'm not a developer, so most of my time goes to reading back what Claude did and checking it did what I asked. The shorter that trail, the sooner I can trust it and move on!

Keep it up...

Let’s go!