we started zarie from a deeply personal need. we've always struggled with executive function: remembering complex supplement schedules, staying in touch with family and friends, and generally keeping on top of the small but important things. we wanted an assistant that lived where we already spent our time and quietly made sure nothing slipped.
we built the first version on telegram. under the hood, it used a multi‑agent architecture: a main agent (zarie) that could invoke specialized worker agents for specific jobs. in practice, that meant we could have a health tracker agent, a social accountability agent, and others, all coordinated through a single, continuous chat. the experience was intentionally single‑threaded so it felt like talking to one real assistant with memory, not a collection of disconnected bots.
on top of that, we built a reminder and automation system for time‑ and search‑based workflows. you could set up things like, "remind me before every india match" or "ping me before every f1 race," and zarie would handle the recurring logic behind the scenes.
we scaled zarie to around 100 users (mostly friends and family) and watched their behavior closely, along with our own. that's when a hard truth became clear: while people liked the idea, it wasn't a hair‑on‑fire problem. the number of genuinely valuable, recurring automations people set up was small, and most users didn't naturally think in terms of creating new workflows. even when we brought zarie into slack, where people already spend their workday, the pattern held. usage was sporadic, not habitual.
at that point, we had to be honest with ourselves. we had built something we personally enjoyed, with an interesting architecture and real users, but not a product with urgent, pull‑based demand.
rather than continue to push a nice‑to‑have assistant uphill, we decided to pause zarie.