Evolving our calendar assistant Reclaim to be AI

The evolution of Reclaim involved rethinking how it handles scheduling to integrate natural language requests while preserving its existing functionality. The need arose because users began describing scheduling goals in natural language, which required connecting these requests with the existing calendar context, user availability, and preferences. Building this integration necessitated a new approach beyond simply adding an AI interface; it required rethinking how conversational input translated into concrete calendar changes while maintaining the core scheduling system.
Engineers addressed this by recognizing that the previous system supported two paths for calendar changes: direct user input or automated scheduling. The emergence of AI agents introduced a third path where users describe goals conversationally. A key challenge was ensuring these agent-driven decisions aligned with the existing scheduling logic, as large language models can interpret requests variably. To achieve this consistency, an agent platform was developed that connects a model to calendar data and Reclaim's features while controlling its actions.
The system relies on a shared internal concept called Schedule Action for all operations, ensuring consistency regardless of the source (user, scheduler, or agent). This consistency is supported by Preview Mode, which allows users to review AI suggestions before applying them, mitigating risks associated with automated changes. The system functions through an agent loop that feeds relevant context and tools to a model, routes actions via Schedule Actions, and employs fast calculation methods in Redis to facilitate real-time previewing.

* Calendars reflect planning, coordination, and reserving time.
* Reclaim is a Dropbox-owned calendar assistant using AI to manage time for tasks, habits, and meetings when inputs are clear.
* Users can describe scheduling goals in natural language that do not map directly to existing settings or commands.
* Connecting natural-language requests requires calendar context, availability, preferences, and commitments.
* The rebuild addressed the need to translate conversational requests into calendar changes while preserving the scheduling system.
* Reclaim previously handled calendar changes via direct user actions or automated scheduling based on preferences.
* AI agents introduced a potential third method for users to interact with Reclaim by describing desired outcomes.
* Variability in large language model interpretation of open-ended requests created challenges for autonomous action.
* The agent platform connects models to calendar details and Reclaim features while controlling access.
* Schedule Action is an internal term for standard operations like creating or updating events.
* Preview Mode provides a temporary view of calendar changes for user review.
* The automated scheduler was reworked as a pure function to calculate schedules without saving to the actual calendar.
* Attendee availability is updated in Redis during preview calculations.

The narrative details a necessary architectural shift from a dual-path scheduling system (manual/automated) to an agent-driven path, focusing on maintaining consistency across interfaces by unifying underlying operations. The core tension lies between the flexibility of natural language input and the rigid constraints of time management involving multiple parties. The solution is not merely adding an interface but fundamentally redesigning the relationship between the AI, calendar logic, and execution.
The decision to unify all scheduling actions under a single "Schedule Action" framework is a move toward system coherence, recognizing that fragmented implementations would lead to maintenance overhead over time. This suggests that architectural consistency is a prerequisite for robust, long-term AI integration within complex domains like time management. The Preview Mode mechanism functions as a necessary safety layer—a mechanism to introduce uncertainty (AI suggestion) into the flow before committing state changes, which reflects a pattern of caution required when automation affects shared resources.
The implication points toward a larger paradigm shift: for AI assistants operating in structured environments, utility is not just about generating novel outputs but about reliably interfacing with existing, interdependent systems. The focus on owning the agent loop and context gathering rather than relying on broad frameworks suggests a resistance to external abstractions that might obscure the specific, high-stakes constraints inherent in scheduling—namely, the consequences of changing time for other people. The freedom gained is framed as the ability to evolve AI capabilities without fragmenting the established scheduling foundation, positioning the system as an expansion layer rather than a replacement.
What systems exist outside this context where consistency across agent interfaces is critical, and what are the hidden costs associated with introducing such safety layers? Does framing the work as "rebuilding" inherently suggest that the previous structure was fundamentally flawed, or merely optimized for a less expressive set of user interactions?