The End of Apps — Kitze, Sizzy.co

Kitze (founder of Sizzy.co and creator of apps Benji, Better, and others) delivers an autobiographical keynote tracing his lifelong obsession with productivity tools and personal agents, arguing that the current era of discrete consumer apps is coming to an end. He frames his talk as a journey from a 10-year-old writing to-do lists in a notebook, through two decades of failed personal productivity experiments, to his present-day exploration of AI personal agents.

Kitze recounts building a series of productivity apps — TodoDoIt (2016, priority-weighted task surfacing), Better (habits + planner + to-dos), and Benji (an all-in-one life OS now spanning dozens of features) — each time abandoning or over-engineering the project due to ADHD-driven feature creep. He notes a key lesson: a competitor who took a single feature from his roadmap — food tracking via phone camera for calorie analysis — made multi-millions while he accumulated 60 features and shipped nothing.

The ChatGPT plugins moment in 2022–2023 (which he correctly identifies as the precursor to MCP) triggered his first major conviction that "apps are over," a belief his wife now dismisses because he calls her with such announcements every few months. He describes the difficulty of coaxing early models to return valid JSON, noting that developers had to "bully the models" to avoid markdown responses. He built a voice-to-API feature in Benji where periodic microphone captures drove live calendar and to-do mutations — a prototype that went viral on Twitter but was never shipped.

Kitze then details his adoption of Claude Code with personal skills, his discovery of the Open Claude (lobster-themed) community and its Telegram/WhatsApp-based personal agent setup, and his subsequent "full hipster mode" — switching to Android for notification access, self-hosting everything on a local NAS, and moving to local markdown files for AI-accessible data. He created multiple specialized Discord bots covering different life domains (work, personal, fitness, etc.) but found the approach produced chaos rather than clarity: missed mortgage payments, late customer emails, and a community that went from explosive growth to five-person "Open Claude Anonymous" meetups.

He identifies core reliability problems: cron jobs failing silently, multi-agent memory loss (agents forgetting context from the message immediately above), and Discord/Telegram being fundamentally the wrong UI for a life OS. He also critiques cloud-hosted agent products (e.g., Cohere's agent) as too neutered compared to self-hosted setups.

Kitze describes building his own personal agent framework, Wolfer, as an experiment (currently only for himself, running on Codex since he fears using Claude Code's REST API). Its key architectural ideas: nested topic trees where each conversation automatically injects parent-topic descriptions as context rather than relying on fallible memory retrieval; visible and collapsible tool-call UI; predictable cron job labeling; per-agent capability management in a sidebar; and an inline knowledge base with markdown documents that can be @-mentioned in any conversation.

His closing prediction is that the role of AI will invert: instead of users prompting AI, the AI will prompt users — surfacing the next task, requesting missing inputs, and operating continuously in the background. He foresees most consumer apps disappearing as OS-level local agents (potentially Apple winning due to strong on-device models and app integration via local tool capabilities) handle normie workflows through dynamically generated UI. Specialist software for domains like video color grading or music production will survive, but general-purpose apps will not.