Loom.

Parallel AI coding agents write code that passes alone and breaks when merged. Agent A changes a function's contract, agent B writes a new caller against the old one, git sees no conflict, and the tests fail only after the merge. Loom keeps the code's call graph, every agent's in-flight edits, function-level leases, decisions and test evidence in one FalkorDB graph. On every edit a CALLS traversal of one to three hops finds exactly the agents whose in-flight code depends on the change, and Loom tells them what changed under them, mid-work. Leases are an atomic Cypher compare-and-set: 50 agents making 10,000 attempts got zero double grants, against 2,332 with the check disabled. Test runs are tied to the body hash they ran against, so a later edit marks a claim stale and blocks it, and any symbol can be walked back through decision, task, agent, change and commit. Measured on the 36 STALE-bench Django task pairs with Claude Haiku, 108 real runs across blind, oracle and live: without Loom 0 of 36 merges passed, with Loom 24 of 36 did, including 12 of 12 signature changes. Most of the remaining failures are return-type changes, where the notice arrived and the model kept its own code anyway.