For example the automated system could take your non-split codebase & inject splits at any async function & self-optimize as needed. The manual annotation support would be to get the ecosystem going as an initial step because I suspect targeting auto-optimizing right from the get go may be too big a pill to swallow.