How to merge parallel AI agent edits safely

Merge parallel agent edits as verified proposals, not as a pile of individually passing patches. Start every worker from the same base, evaluate each edit alone, identify interactions through shared acceptance conditions, then test the selected combination. If the dependency map or evaluator is incomplete, treat the result as a review candidate rather than a certified safe merge.

The conflict is in behavior, not necessarily in the diff

The paper's example has a pricing module and a checkout module. Either can apply a required discount. Each agent's edit improves the same starting artifact, but applying both discounts is wrong. A separate receipt edit does not share that acceptance condition and can be combined with either. Thus a clean text merge, or even edits to different files, does not prove behavioral compatibility.

Model a dependency edge when two edited blocks can affect the same objective factor: a test, interface contract, invariant, or decision rubric. The paper notes that extra edges make the certificate conservative but reduce parallelism; a missing real edge can incorrectly label interacting edits independent. For the upstream handoff, validate the cross-agent coding specification before workers start.

Run a four-step integration gate

This is a practical adaptation of the paper's proposal-integration model, not a claim that a typical CI suite is an exact objective. Where the objective really decomposes over known factors, an independent set of the dependency graph has additive verified gains. The authors prove a worst-case retained-gain bound involving the graph's fractional chromatic number for their fixed-proposal model. A CI result with untested behavior cannot inherit that bound.

Git's branching and merging guidance addresses integrating branch histories and resolving merge conflicts. Use that mechanical merge process, but run behavioral acceptance on the resulting tree; absence of a textual conflict is a different claim.

Read the experiment as a scoped signal

In the paper's seven three-module repair tasks, the authors identify cases where two adjacent edits each pass an interface check alone but fail together. Across the first-round conflict subgroup, naive merge succeeded in none of 11 runs; the verified integration methods succeeded in all 11. Those are small, constructed repair tasks, not a measured production defect rate. Route-scheduling tests reported a median 15% wall-time reduction across 12 paired runs, while route-cost changes varied. Parallelism also raised input-token-equivalent cost on larger conflict-free tasks.

For a real repository, keep a serial baseline and measure elapsed time, evaluator cost, token cost, and accepted behavior together. If the combined edit fails, preserve each isolated proposal and the failed combined test so the next worker can diagnose the interaction. See agent failure root-cause analysis for tracing the first consequential failure, and when multiple agents earn their coordination cost for the broader architecture choice.