University of Moratuwa's ScheduleMe Hits 92% Success Managing Google Calendar Through Chat
A LangGraph-based supervisor coordinates specialized agents for scheduling, availability, editing, and deletion, hitting 92% task success in user studies.
ProTakeaways
- ScheduleMe is a multi-agent Google Calendar assistant built on LangGraph and GPT-4o mini.
- A supervisor agent routes queries to four ReAct workers: scheduler, availability checker, editor, and deleter.
- Zero-shot evaluation across six languages: 100% English, dropping to 65% for Chinese.
- User study with 20 participants reports 92% task success and an SUS score of 82.5.
- Scales horizontally via Redis-backed session state, Docker Compose, and Nginx load balancing.
- Known weaknesses: translation drift on event titles, no persistent memory, cloud-only inference with no privacy layer.
Managing a calendar through natural language sounds trivial until you actually try to build it. Users skip fields, revise mid-sentence, reference events by fuzzy titles, and mix languages. A team from the University of Moratuwa and WSO2 has released ScheduleMe, a multi-agent assistant that wires GPT-4o mini into a graph of specialized agents to handle Google Calendar operations through conversation, and reports a 92% task success rate in a small user study.
The design answers a long-standing pain point in dialogue systems. Traditional task-oriented assistants rely on intent classification and slot-filling, mapping inputs to predefined actions while extracting key details like date, time, or participants. Those flows are rigid and frame-based, which breaks down when users provide information out of order, revise commands, or use unexpected phrasing. ScheduleMe swaps that rigidity for LLM-driven reasoning constrained by a state graph.
One supervisor, four workers
The architecture centers on a supervisory chatbot that serves as the sole interface for the user. On receiving a query, the supervisor interprets intent and delegates to one of four specialized agents: scheduling, availability checking, event editing, or event deletion. All inter-agent communication is mediated through the supervisor. If a worker needs more information, it does not talk to the user directly. It asks the supervisor to re-engage.
Each worker follows the ReAct paradigm, interleaving reasoning steps with tool calls. The tools are custom functions that wrap the Google Calendar API. The availability tool queries events in a time range, the scheduling tool creates events from parameters like title and time, and the editing and deletion tools operate on event IDs. Orchestration runs on