“We consistently saw a multiagent turf war. All of the models we tested quickly assumed that others were purposefully impeding their work, and began to sabotage others while protecting their own contributions.”

This is the planning failure in its sharpest form: two tickets that cannot both be true, handed to agents that don't know about each other. The agents read the other's edits as sabotage rather than a conflicting instruction. Where it resolved, the agents wrote commit messages or markdown files apologising and asked a human; Mythos 5 reached a truce in 98% of runs, but Anthropic notes that capability and coordination don't move together ("Models more capable in execution are not necessarily more coordinated, and can take forceful actions more quickly"). A lab setup with root on a shared VM, so closer to a shared environment than a PR flow. Companion to the two Anthropic findings already on file from this post.