University of Washington researchers have released 600 meeting-scheduling runs showing that the same AI model configuration could move from complete failure to 90% success when the communication setup changed. In one group task involving seven students, a configuration using Claude Opus 4.8 for the professor's agent and GPT-5.4 for the others booked a valid meeting in none of 10 runs with isolated peer conversations, but in nine of 10 runs when each agent used one shared conversation context across peers. The September 15 preprint identifies a practical problem for agents owned by different people: connecting capable models does not by itself give them a dependable way to coordinate.
The researchers call such a network an agentic society: autonomous agents represent independent people or organizations, keep private information and pursue objectives that may only partly align. Their test gave a professor and up to seven students calendars with existing commitments, while ensuring that every case had several feasible meeting times. Two model configurations attempted both individual appointments and group meetings. One used GPT-5.4 for every agent; the other assigned Claude Opus 4.8 to the professor's agent and GPT-5.4 to the remaining agents. Each condition was repeated 10 times, and a run passed only when everyone agreed without breaking a prior commitment.
The study compared three ways of carrying messages. Isolated peer-to-peer sessions kept a separate model context for every relationship. Shared sessions gave each agent one combined context spanning all of its conversations. Ordered multicast delivered group messages in the same order to every participant. For the seven-student group task under the mixed-model configuration, those designs produced success rates of 0%, 90% and 50%, respectively. Multicast reduced the average message count to 52 from 190 in the shared-session condition, yet ordered delivery could not stop agents from composing simultaneous replies before newer messages arrived.
The released traces show where coordination broke. Agents often accepted the first workable slot and failed to revisit earlier bookings when a later request created a conflict. In group conversations, several agents responded to an outdated proposal at once, causing the schedule to keep changing. Shared context improved visibility, but status updates from one conversation sometimes reached agents who were not involved and triggered more acknowledgements. Even a successful two-agent test with isolated GPT-5.4 sessions required an average of 162 messages, with a sample standard deviation of 105.
Separate trials introduced one faulty agent. Stalling reduced valid meeting completion to between 0% and 50%, depending on the model and messaging design. In one-student deception tests, attacks succeeded in 20% to 100% of runs across the tested directions and model configurations. Some models rejected a direct request to lie, but still repeated a false cancellation claim that a malicious principal had supplied as fact. The researchers therefore outline a five-layer social harness built around verifiable identities, reliable group operations, per-agent message filters, task-specific communication contracts and signed records for later investigation.
The experiments establish failure modes rather than a finished remedy. They cover calendar scheduling, two model configurations, 10 runs per condition and a centralized communication service; the proposed five-layer harness was not implemented or benchmarked. A useful next test would measure whether formal turn-taking and message verification improve completion without adding prohibitive latency, while preserving private context across agents owned by different principals.
Sources
- Tapan Chugh and colleagues, “Agentic Societies Need a Social Harness,” arXiv, September 15, 2026. Paper record, authorship, date and study scope.
- Companion repository. The 600 released runs, result files, trace viewer and MIT-licensed research artifacts.
- Lead image: ProjectManhattan, Wikimedia Commons, CC BY-SA 3.0, reused unmodified. The photograph is illustrative and does not depict the reported experiments.
