Follow the Text: Channel-Agnostic Detection of Inter-Agent Communication via Provenance

We present a channel-agnostic detector for such communication that does not require identifying communication channels ahead of time. We tag every string read by an agent with its origin, and flag instances that were authored or relayed by other agents. On published logs of a reproduction of the HF incident, 93–99% of the communication edges we detect are independently backed by a prior registry write from the origin agent. In an experimental rig with file-system ground truth, our detector raises one false finding in 20 isolated episodes, finds every flow in the positive control and in authored messages, and 0.91–0.92 of relayed handoffs on Claude Haiku 5.5 and Sonnet 5.5. On refactoring tasks from Refactor Arena it raises 55 false findings per 1,000 inputs. Models briefed on the monitor declined to evade it, but a scripted white-box adversary evades it with four of six strategies, all of which keep the payload from crossing a model boundary as the same text. The backing store for our detector provides scan time linear in text, with no pairwise agent comparisons. We project that it would require 50 bytes per token in a native implementation, making it feasible to use at scale.