Sonifying multi-agent AI telemetry so drift and failures are audible, not just logged

Built OtelJazz after getting frustrated watching multiple agent logs scroll past in parallel and not actually tracking any of it. A trace tree gives you zero coverage while you're doing anything else, reading code, in a meeting. Hearing is peripheral and continuous in a way a log window isn't. The mapping: seven piano voices fuse into one chord that thickens and thins as agents spawn and finish. A voice sliding off the shared rhythmic grid while the others stay locked means an agent diverging from its mandate. Voices that should be independent locking into unison is a candidate collusion signature. A voice dropping out mid-phrase is a silent stall. Tempo tracks span throughput. Design choice that mattered most: telemetry only drives rhythm and density, never harmony. Tried letting telemetry drive chord changes too, early on, and it made the whole thing indistinguishable from noise, nothing stable to hear the dynamics against. Scoping honestly: the demo plays a synthetic swarm generated in your browser, not a live feed, though real capture exists in the repo (a hook writing actual gen_ai spans, an OTLP receiver) and just isn't wired to the page yet. No listening study has been run, so whether anyone can actually decode this by ear under real working conditions is genuinely open, not a claim I'm making. Press play, no signup. Code's MIT