DNS for Human Time

It started with a simple observation: when you ask an AI agent to schedule a meeting "next Tuesday at 2pm," it gets the date wrong about half the time. The OOLONG benchmark confirms this — even frontier models score below 50% on temporal reasoning tasks.

We looked at the existing calendar tools for AI agents. They were all thin CRUD wrappers — list events, create events, delete events. They delegated all temporal reasoning to the LLM. No locking, no conflict detection, no availability computation. They weren't solving the problem; they were passing it through.

The insight was that calendar scheduling is not a language problem. Resolving "next Tuesday" to a date is not ambiguous — it's arithmetic relative to today. Expanding a biweekly RRULE across a DST transition is not judgment — it's an algorithm defined by RFC 5545. Booking a slot without double-booking is not coordination — it's a lock.

So we built the computation layer that LLMs are missing: the Truth Engine for deterministic RRULE expansion, Two-Phase Commit for conflict-free booking, and a cross-provider availability graph that merges Google Calendar, Outlook, and iCloud into a single view.

The LLM does what it's good at — understanding intent. The tools do what they're good at — computing the answer. That's the division of labor that makes AI scheduling reliable.