AINEWS
Git was built for one developer at a time. Multi-agent coding is breaking that.
When two AI coding assistants edit the same file, git's old line by line merge gives up. A new wave of entity level tools parses the code's structure first.
When two AI coding assistants edit the same file, git's old line by line merge gives up. A new wave of entity level tools parses the code's structure first.
Two AI coding assistants open the same source file on a Tuesday afternoon. They work on different functions. Their changed lines do not overlap. Git still flags a conflict, because the only thing it knows how to compare is text.
That failure is not a bug. It is a 2005-era assumption baked into git's merge logic: that a file is a stream of lines, and a conflict is two streams disagreeing. The assumption is starting to crack now that multiple AI agents routinely edit the same repository at the same time.
Weave, a new open-source merge driver from Ataraxy Labs, is one of the more concrete attempts to replace that assumption. Instead of asking git to reconcile overlapping lines, Weave parses the code with tree-sitter, a parser that turns source files into a structured tree of functions, classes, and methods, and merges at that level. If two agents edit different functions in the same file, there is no conflict to resolve at all.
The project describes itself as an "entity-level semantic merge driver." In practice, that means three layers. The Merge Driver slots into git, replacing git's built-in line-based merge step, which today does not understand what a "function" is. Coordinate, built on a CRDT (a data structure, originally designed for collaborative document editors, that lets multiple writers make claims about a shared resource without a central server arbitrating), lets agents claim an entity, such as a function or class, before they start editing, so other agents know to stay out. Connect exposes the whole system to assistants like Claude through the Model Context Protocol, a relatively new standard that lets a language model call external tools as if they were local functions. Connect ships 15 such tools.
The hard question is whether this is a real category shift in version control, or a tidy product story on top of a familiar problem. The Weave project page reports a benchmark of 31 out of 31 merge scenarios across seven languages, compared with 26 out of 31 for mergiraf, a syntax-aware rival, and 15 out of 31 for stock git. The same page claims 83 wins and 0 regressions across 4,917 file merges in C, Python, and Go, with more than 1,500 downloads. None of those numbers have been independently verified in the current source set, and the benchmark methodology is not visible on the page. Treat them as the maker's claims, not as settled fact.
That caveat matters because the underlying idea is bigger than any one tool. Once AI agents routinely write code in parallel, the unit of a "diff" stops being a line and starts being a function, a class, or a named entity. Syntax-aware merging has been a research topic for years. What is new is the demand for it: multi-agent workflows that produce, in a single afternoon, more concurrent edits to one file than a human team used to produce in a week.
There is also a coordination layer to watch. CRDTs are the same data-structure trick that lets Google Docs and Figma handle concurrent edits without a server deciding who wins. Applying the same idea to code entities is plausible on paper and unproven in production. The Model Context Protocol side, where the merge tools become a callable service for an AI assistant, is even newer: it only works if the model can be trusted to claim the right entity before it starts typing, and to notice when it has not.
For a developer already drowning in merge conflicts from a Copilot-style pair-programmer, the practical bet is small. Weave installs with brew install weave, drops into git as a configured merge driver, and stays out of the way when the file is small or the change is simple. For a team running multiple agents in parallel, the bet is larger: that the future of "merge" is less about reconciling text and more about agreeing, ahead of time, on which piece of the code each agent owns.
Either way, the line is moving. The merge conflict was, for two decades, a problem about text. It is now, slowly, a problem about meaning.