Databricks wants AI agents to take their working state along on experimental detours. In a November 20 post, Databricks’ @ankit_math said the team had previewed its new agent platform at AI Engineer Code Summit. The platform is in development, with proposed branching support for sandbox, storage and memory. Source
“What if you could git branch your entire agent?” the post asks. The proposed workflow is to fork that state, try something risky, then throw the branch away or merge it back. Source
The organizer’s account frames the motivation around the files an agent uses: “Your agent branches. Its files need to branch with it.” It specifically names notes, plans and scripts as things each investigation needs its own copy of, describing the session as being about “branching the entire working state.” Even the notes are part of the experiment. Source
In a November 21 quote-post, @alighodsi described the scope as “Code, sandbox, database.” The post doesn’t explain how database branching would work. Source
The posts don’t provide a product name, availability date or rules for merging state back together. The Git comparison doesn’t establish Git compatibility, either. Branching gets the catchy analogy; merging still needs the explanation. SourceSource
Comments