Software assets in the AI era

Thanks to agentic engineering, the cost of code is getting closer and closer to zero. Even in such an era, what software assets still retain value?

Source code is a derivable output

Since source code can be compiled into a binary file, it is usually treated as an important asset. That is why there is an open source movement, but no open binary movement. (But does one really not exist? Maybe if I look for it, there might be one.)

But now, if you have a well-made test suite, producing source code is becoming easier and easier. From the perspective of BDD, a test suite is an executable specification. In other words, if you have a well-written executable specification, source code is becoming almost free.

Are tests and specifications valuable?

Then are tests valuable outputs?

It does not seem so. Because while it is possible to create source code from tests, it is also possible to create tests from source code.

Of course, there is a bit of a difference in that the tests created this way capture not “how the system ought to behave (that is, the specification)” but rather “how the system currently behaves (including bugs).” In Working effectively with legacy code, this kind of test is called a Characterization test. Tests tend to contain a bit more of the “intent” that could not be fully captured in code. They seem more valuable than code.

Meanwhile, if you decompile a binary file, you can obtain source code. There are problems such as comments not remaining and many symbols not being recovered, but thanks to AI, making it readable has also become somewhat easier. In other words, if you have only a binary file, you can obtain source code, and if you have source code, you can obtain Characterization tests, so if you have any one of the three, it has become relatively easy to obtain the other two. (Or rather, that is what it is becoming. As of 2026-03-01.)

Then what remains?

It seems that what remains are intentions not fully captured in the specification, the trajectories of decision-making and execution, and the creator’s tastes and user feedback behind those decisions and intentions (VoC, behavioral data, etc.).

As expected, taste matters. (See also: Software imbued with taste)