The more I use Claude Code, the more impressed I become.
I keep throwing progressively harder problems at it, and whenever the problem is conceptually tractable, it can usually just solve it. Not with hacks or brittle workarounds, but by actually engaging with the structure of the problem.
I decided to push it further by combining two difficult problems.
Difficult in the sense that either one would likely take me years to complete properly on my own let alone together whether the interaction between the two is a kind of complexity in its own right.
With the exception of one genuinely catastrophic error that required intervention to recover from, the tool has kept going, iterating, and making real progress.
What stands out most is that it seems to understand what progress actually is.
It does not treat the number of passing tests as a sacred metric. It is willing to break tests if that moves the system forward in a deeper, more honest way. That is something many humans struggle with.
The mere fact that it can reason about progress at all, rather than optimising a superficial proxy for it, is pretty remarkable
And to think the tool over past few months has seen pretty consistent improvement at the cadence of weeks with no end in sight.
Will software development be unrecognisable a year or even 6 months from now, I do not know.