Ask ChatGPT to analyze a spreadsheet, and it might write and run a Python script to find the answer. Once you have the answer, you may never need that script again. That is just-in-time, disposable software, and I expect most software run in the cloud will soon look like it.
Today, we build applications around needs we can anticipate. We invest in code review, testing, and deployment pipelines because we expect people to keep using what we ship. When writing a program costs almost nothing and takes seconds, building one to answer a single question starts to make sense.
Source control matters less for programs nobody will revise. We’ll need verification that works as code is generated and executed, without a human reading every line. Permissions and spending limits have to hold even when the program was written moments before it ran, and we’ll still need a record of what it did.
Agents will also take over the infrastructure decisions engineers now make in tools like Terraform. One task needs a few seconds of compute; the next needs a powerful machine for days. The agent will acquire those resources through cloud APIs, within the task’s budget and security constraints, and release them when it’s done. This is application-defined compute: the software chooses its infrastructure as it does the work.
This will make it practical to solve problems that never justified a full software project. A task can have its own software, even if nobody ever needs to do it again.
