Six agents. One file. Same moment. Chorus merges simultaneous edits per symbol, and the converged file is a pure function of the intent set — any land order, same bytes.
Chorus is the merge engine under Loom's fleet: it lets many coding agents edit the same file at the same moment — no locks, no turns. It merges per function rather than per line, surfaces the one edit pair that truly collides instead of guessing, and converges to byte-identical output in any land order. Four acts, ~50 seconds, then it loops.
The complete story for someone who has never heard of a merge conflict: why six agents on one file breaks everything, what the industry has tried — locks, git, file ownership, CRDTs, LLM merging — where each one falls short, and how Chorus puts members, visible checks, and order-independence together. Eight chapters.
Three wireframes from the paper. Every claim below is true of the shipping engine — the figures are redrawn from the paper's own diagrams.
Files become symbol cells, edits become intent. tree-sitter decomposes every file (12 languages) into per-symbol cells plus fenced "other" regions. Agents never submit diffs — they submit Edit / Add / Remove ops keyed by cell, and the merged file is recomputed from the fixed origin plus the accumulated set.
Any arrival order, the identical file. Lands accumulate raw intent into an order-free set; the converged (canonical, conflict-set) is recomputed from scratch as a pure function of that set. Permutation invariance isn't tested in — it's true by construction, then cross-checked on every fuzz case.
Structure isn't enough — behavior is gated too. After the structural merge, C2 replays every writer's own green tests on the merged tree. A test that was green on a writer's tree and is not passing on M is a silent loss: the merge surfaces instead of shipping. Absent or crashed suites count as failures — the gate fails closed.
Fleets of AI coding agents make concurrent edits to the same file the common case rather than the exception. Merge queues restore safety by serializing — surrendering the parallelism the fleet exists for — while line-based three-way merge resolves concurrent edits order-dependently and without language structure. Chorus takes a third path: files decompose into per-symbol cells, writers submit intent rather than diffs, and the converged file is recomputed as a pure function of the fixed origin and the accumulated intent set, making land-order independence a structural property rather than a test target.
Real conflicts do not disappear: same-cell contention, removals that would dangle a surviving reference, and every ambiguous construct fail closed and surface per cell while the rest of the file lands. A semantic gate then replays each writer's own passing tests on the merged tree and blocks silent behavioral loss. We evaluate with sixteen rounds of adversarial review that produced mechanically-reproduced counterexamples, and a mutation-kill audit of our own fuzzing oracle that found — and fixed — three blindness classes the campaign could not see.
Open the PDF@misc{beckmann2026chorus,
author = {Beckmann, Yiming},
title = {Chorus: Order-Independent
Structural Merging for Fleets
of Concurrent AI Coding Agents},
year = {2026},
note = {MingLLM Technical Report},
url = {https://loom.mingllm.com/chorus}
}