← back
Viktor: AI Coworker That Lives in Slack — Fryderyk Wiatrowski
Takeaway
Building a company-level AI coworker in Slack hinges on permissioned multi-user memory and collapsing Slack's many interaction modes into a coherent agent context.
Summary
- Fryderyk Wiatrowski (Viktor co-founder) pitches Viktor as an AI employee that lives in Slack with access to 3,000 integrations, inheriting permissions across the company instead of per-user connections.
- Company history: started as JCE (web agent) in 2023 — pre-tool-calling, used minified DOM snapshots (state of the art on WebArena) but only reliably handled 3-5 steps at ~60%; pivoted to Jayce (email agent on Sonnet 3.5) using face-first agent loop; Viktor launched Feb 2025.
- Slack chosen because (1) you don't talk to human coworkers in web apps and (2) 10-minute task latency feels normal in Slack but unbearable in a chatbot interface.
- Memory challenge multiplied: open-claw-style memory clutter scales 100x with multi-user; Viktor solves cross-channel context isolation (growth channel context must not leak to engineering or DMs).
- Slack-as-input is multi-modal: DMs, threads, public channels, emoji reactions, edits, deletes — must collapse into linear agent context (delete = stop task, edit = re-respond, forgotten thread = sandbox decision).
agentsslackai-employee
Original description
Viktor is an AI employee that lives in Slack. No web UI. It participates in channels and threads the way a teammate does, inherits integrations from whoever connected them first, and handles tasks that take ten minutes while you move on to something else. This talk covers what breaks when you scale a personal agent to a whole company. Slack is a more complex input surface than it looks: threads, DMs, edits, deletions, emoji reactions, and conversations that drift between channels. Memory isolation gets harder when the same agent needs context for a hundred users without leaking the growth channel into the engineering queue or one person's DMs into the team feed. And when you try to swap the underlying model for something cheaper, users notice in ways that have nothing to do with task performance. Speaker info: - https://x.com/fawiatrowski - http://getviktor.com/