Inherit Shared Integrations for Horizontal Company Context

Company agents like Viktor differ from personal ones (e.g., OpenClaw) by requiring only one user to connect integrations—Viktor then inherits permissions across the team, accessing 3,000 tools like Posthog analytics or Meta ads without repeated setups. This avoids the pain of 20 growth team members each linking their own Meta account, reducing errors from mismatched integrations. Result: broad, PhD-level context spanning codebases, emails, and analytics, enabling cross-role tasks a specialized CMO couldn't match. Trade-off: accidental oversharing (e.g., one user's personal Gmail exposed team-wide), solved by scoping integrations to DMs or specific users, mimicking how you wouldn't give a new hire your personal email.

Slack's interface beats web apps for long tasks (10+ minutes) because users expect delays—like pinging a teammate for an app build—lowering frustration versus waiting in a chat UI tuned for 30-second responses. Viktor runs cloud-based, no desktop needed, unlike desktop agents like Cursor.

Isolate Memory and Manage Slack's Messy Inputs

Scaling from one user to 100x memory load demands isolation: growth channel context stays out of engineering or support feeds, DMs don't leak team instructions unless the user belongs there. Conflicting instructions across hierarchies require structured access rules. Slack's inputs explode complexity beyond linear threads—DMs, channel threads, edits, deletions, emoji reactions, and drifts (humans abandon threads for new DMs to the same person). Solutions: roll over context from prior DMs/threads into new ones; halt tasks on deletions; respond to edits; linearize all into agent context without leaks.

Early browser agents (JCAI, 2023) handled DOM snapshots for universal access but failed after 3-5 steps at 60% reliability due to compounding errors and minute-long waits—state-of-the-art on WebArena benchmark but not product-ready. Email agent Jayce improved with Sonnet 3.5, triggering loops on arrivals for drafts or gated tool calls like refunds.

Model Personality Trumps Raw Performance; Earn Proactivity

Users detect model swaps instantly: GPT-4o excels at tool-calling/codegen and costs less than o1-preview (Opus), but lacks o1's sassy personality—AB tests sparked user backlash. Viktor's architecture amplifies this; stick with what fits the tone for retention.

Proactivity shines when earned: Viktor scans Posthog mid-AB test discussion, validates claims (e.g., flags non-significant results with calculations), or suggests automations. Day-one aggression (DMing everyone) alarms security; start with few users, prove value, then expand. Three pillars to build your own: (1) Get work done via models + Pipedream connectors; (2) Know the company by leveraging full Slack context (post-Slack approval); (3) Make it friendly—personality drives adoption.

Vision: Every company hires AI employees for all cognitive tasks, echoing Leibniz's 17th-century push to automate calculation drudgery.