“Split a feature between two coding agents and each one passes its own tests. Merge their branches and the build is red. One agent changed a function. The other, on its own branch, was still calling the old version. With work trees alone, we measured zero of 18 merges green.”

Novel mechanism: symbol-level coordination enforced by a pre-edit hook (a graph query, not a model call), plus memory attached to code rather than sessions. Author's own small-pair benchmark, so treat the numbers as vendor claims. Telling each agent the other's task recovered only 5 of 9, which is evidence against prompt-level awareness. Cost: held agents wait.