Paper Tests Claude Opus 4.8 Agents on Conflicting Code Edits

The paper studies a scenario that any team running several coding agents against one repository will recognize: two separate, independently motivated requests arrive at once, each agent solves its own task, and the two solutions have to be merged into something that still works. The authors call this setting “multi-agent by necessity,” since nobody designed it that way, it just happens whenever more than one request lands on a shared codebase at the same time.

To measure it, the paper introduces ParallACT, a benchmark built from two kinds of task pairs. Complementary pairs are tasks that are each separately useful but need a shared interface to combine cleanly. Contradictory pairs require editing the same code in incompatible ways, so a naive merge of two independently correct solutions produces something that satisfies only one of the original requests. The researchers compare a single oracle agent handed both tasks together against several multi-agent arrangements that vary execution order, whether agents can see each other’s work, whether they are told a peer exists, and whether they share a communication channel.

The headline result is a sharp drop for parallel agents relative to the oracle, worst on contradictory tasks. For Claude Code running Claude Opus 4.8, the contradictory-task score falls from 62% for the oracle to 16% once two agents edit independently and their changes are merged. The paper reports the same pattern across six further pairings of two agent harnesses with three open-weight models, which suggests the problem is not specific to one vendor’s agent but a structural weakness in running agents without coordination.

Two mitigations tested barely help. Simply telling an agent that a peer is working on the same code does not meaningfully change outcomes. Giving agents a communication channel recovers only a small part of the lost performance, and it comes at an additional token cost, meaning teams would pay more per run for a partial fix rather than a reliable one.

The numbers

The 62% to 16% drop on contradictory tasks is the clearest figure in the paper for anyone deciding how to structure a multi-agent coding workflow today. It implies that where two agents’ tasks are likely to touch the same lines, serializing the work, or routing both through a single agent, still beats running them in parallel and hoping a merge step or a chat channel will catch the conflict. The paper frames this as a benchmark meant to isolate exactly which coordination strategies help and at what cost, rather than a finished solution, which leaves the practical choice of serialize, merge carefully, or pay for communication squarely with the team operating the agents.