Viktor: AI Coworker Scaled for Slack Teams

Agent Evolution to Company Scale

Viktor started as an experiment in February 2024 after browser-based JCAI (3-5 steps at 60% reliability on WebArena benchmark) and email agent Jayce (triggers on arrivals, drafts replies or calls tools like refunds with approvals). Viktor shifts to company-wide: one user connects integrations (e.g., Meta Ads, PostHog), whole team inherits access without reconnections. It lives solely in Slack—no web UI—participating in channels, threads, DMs like a teammate, using broad company context across roles (e.g., CMO-level with codebase access).

Slack Input Challenges and Fixes

Slack's surface includes threads, DMs, edits, deletions, emoji reactions, channel drifts. Agents linearize non-linear inputs: on message delete, halt tasks; on edits, respond accordingly; on new DMs mid-thread, roll over prior context to avoid new sandboxes. Memory isolates per user/channel (e.g., growth channel context stays out of engineering or DMs unless relevant). Proactivity earned gradually: starts passive, then suggests automations (e.g., joins AB test discussion, queries PostHog, flags non-significant results with calculations) to avoid security backlash.

Personality and Model Choices

Users detect model swaps beyond task performance. A/B test swapped Opus (current main model, sassy tone) for GPT-4o (stronger/cheaper tool-calling/codegen): users raged, preferred Opus personality. Integrations scoped: shared by default (company tools), personal optional (e.g., Gmail in DMs only). Viktor as 'hire,' not tool—avoids personal data leaks like one e-com customer's Gmail incident.

Pillars for Building AI Coworkers

1. Helps get work done: models capable today; connect via Pipedream (3000+ integrations).
2. Knows company: utilize Slack context fully (post-Slack approval).
3. Friendly: tune personality so team likes agent. Vision: every company hires AI employees for all cognitive tasks, echoing Leibniz's 17th-century calculator push.