What's this about?
Joint generation and maintenance of multiple applications beyond the single codebase. 24 upvotes.
Why is "beyond the single codebase" the hard part?
Because almost all coding-agent work assumes one repository with one owner. Real software is a portfolio: several applications sharing libraries, where a change to the shared code must be validated against every consumer, and where the consumers have different owners, release schedules and test suites. This is the problem that makes large organisations slow, and it is completely absent from the benchmarks agents are measured on. Generating one application is a solved-ish problem; keeping five consistent as they evolve is not.
Why does this matter to you?
- Maintenance across repositories is where the actual cost of software lives, and almost nothing measures agent performance on it.
- It is the multi-repo version of the SWE Refactor Bench argument from last week — whole-system consistency, not per-file correctness.
- If you are running agents across more than one repository, this is the failure mode you have already hit.