Every day I have multiple worktrees and subagents making progress on the same task, and they continuously overwrite or conflict with each other’s critical Convex functions. Giving every agent its own development account is not workable: they are collaborating on the same task and need access to the same underlying data. If my agents have multiple worktrees, each npm dev crashes the code of another agent, and so anything net new can crash the other agent. Additionally even on the same branch, the other agent doesn't know what the current agent changed and gets confused.
I’d like first-class, functional database forking for agentic development: each worktree/agent can run or request an isolated code branch of what they are seeing in the shared data, so agents can safely change and test functions without stepping on one another, while the data model remains shared and application code stays agnostic to the particular fork. Changes could then be deliberately merged or promoted once they are ready.
A another solution could treat a development fork as a versioned combination of database data and functions, with explicit merge and synchronization controls for both. That could make the current faux-agent-dev site split viable: agents could stay on separate function branches while safely syncing the shared data they actually need, resolving conflicts intentionally instead of colliding at runtime. I would prefer the previous solution.