Tarmac, a project published on Alibaba Cloud's community blog, models airline disruption recovery as a negotiation between five Qwen agents: rebooking, crew legality, gates and ground, hotels, and a passenger advocate. Each places revocable claims on a row-locked seat ledger; contested seats go to sealed-bid rounds so no agent can watch another's moves; deadlocks are resolved by a duty-manager mediator running a larger Qwen model, which issues Ed25519-signed rulings citing the rule they rest on. The reported benchmark uses one frozen scenario, 180 displaced passengers against 139 legal seats, run across 10 random seeds. Against a single optimisation planner, the agent society stranded 3 protected passengers versus 17, met 100% of special-needs service levels versus 0%, and committed no crew-duty violations, according to the authors. The implementation splits models by role: the five negotiating agents run on qwen3.7-plus through an OpenAI-compatible endpoint, the mediator on qwen3.7-max. Every step is written to a hash-chained log that replays byte for byte; the repository includes a live model run with 324 log entries and a test suite of 330 tests. The orchestrator is deployed on Function Compute and the code is MIT-licensed.
Our takeLegibility, not the headline score, is the claim worth testing: five named interests argue on the record, so a trade-off can be inspected instead of inferred from an objective function. Two limits remain. The result rests on one frozen scenario across 10 seeds, a small sample for a stochastic negotiation, and the regulation the mediator cites is whatever the scenario supplies, so ambiguous real rules stay untested. The hash-chained, replayable log is the more reusable idea: if agent decisions must be auditable, log structure matters more than model choice.