Korva is a scroll of the internet’s best ideas.

Picture a five-step function written in TypeScript and running on AWS. Because its state is stored after each step, the function can be rewritten in Python and run on GCP. The platform calls it there instead, hands it everything from the prior three steps, and it finishes the last two in the new language.

That makes compute fungible, and possibly programming languages too. A function written in Go can likewise be rewritten in TypeScript and continue from where it left off.

The same property helps with AI. Each LLM call is a step and each tool call is a step, so the run happens in an environment where the same input replays the same way. That determinism came from the original design of the system, which sits between application code and the infrastructure, and it carried over to agents with no changes.

How can a function resume in another language?

The state saved after each step is plain data, such as JSON, rather than a language's own memory. When the function runs again, the code starts from the top, and every step that already finished returns its stored result instead of running. A function in a different language follows the same path through the same recorded results. Nearly every programming language ships with a JSON parser in its standard tooling.

What makes a replayed function repeat itself?

The code between steps has to reach the same decisions when given the same inputs. Sources of change, such as the clock, random numbers, and network calls, are wrapped inside steps so their results get recorded and reused on a replay. Workflow engines like Temporal impose the same rule, and its documentation requires workflow code to be deterministic.

Why does movable compute matter to clouds?

Prices and capacity differ between cloud providers and between regions, and work that is not tied to a machine can go wherever it is cheapest or available. Spot instances show why saved progress is valuable: AWS can reclaim a spot instance with a two-minute warning, and jobs that checkpoint their progress can carry on elsewhere.

More from The Peel with Turner Novak

Tech & AIMonths to build, a day with stepsA healthcare company's basic patient flows took months to build, and step functions that wait for an event would have cut that to a day or two.
Enjoyed it?