Viktor: AI Coworker That Lives in Slack — Fryderyk Wiatrowski

Key Concepts

• AI Employee: An autonomous agent that functions as a team member, possessing company-wide context and tool access.
• Slack-Native Interface: Using Slack as the primary UI for AI agents to mimic human interaction and manage long-running tasks.
• Shared Context: The ability for one team member to connect an integration (e.g., Meta Ads, PostHog) that becomes accessible to the entire team.
• Proactivity: The agent’s ability to monitor discussions and offer insights or execute tasks without being explicitly prompted.
• Agentic Reliability: The challenge of maintaining high success rates in multi-step workflows.

---

1. Evolution of Victor

Victor began as an experiment in February 2024, evolving from earlier iterations of the company's technology:

• JCAI (2023): A web-based agent that operated by taking snapshots of the DOM (Document Object Model) to navigate websites. It faced significant reliability issues (approx. 60% success rate) and latency problems.
• Jace: An email-based agent that triggered workflows based on incoming emails (e.g., automated refunds).
• Victor: The current "AI Employee" that lives in Slack, integrating with over 3,000 tools to provide horizontal, company-wide context.

2. Technical Architecture and Frameworks

• Integration Strategy: Victor utilizes a "shared context" model. Unlike personal agents where every user must connect their own tools, Victor allows one user to connect an integration, which then inherits permissions for the whole team.
• Slack as the Operating System: The team chose Slack over a web app for two reasons:
  1. Human-like Interaction: It feels natural to collaborate with an AI in a channel or DM, just like a human teammate.
  2. Latency Management: Complex tasks that take 10+ minutes to execute are perceived as "normal" in Slack, whereas they are frustrating in a web app where the user expects immediate responses.
• Context Management: The system must handle complex Slack inputs, including message edits, deletions, and thread-switching. The agent must maintain a "linear context" even when users start new DMs or switch channels.

3. Key Challenges and Solutions

• Memory Management: Scaling an agent from one user to 100+ users creates a risk of memory clutter. The team implemented architectural solutions to prevent memory exhaustion.
• Security and Access Control: A major challenge is preventing context leakage (e.g., ensuring growth team data isn't leaked to the engineering channel). The team developed granular scoping for integrations to ensure users only grant access to what is necessary.
• Personality and Tone: The team discovered that model "personality" significantly impacts user adoption. They found that users preferred the "sassy" nature of Claude 3.5 Opus over other models, even when those models were technically capable of the same tasks.

4. Real-World Applications

• Analytics Monitoring: Victor can monitor tools like PostHog during team discussions. If a team discusses an A/B test, Victor can proactively verify if the results are statistically significant and provide a calculation to support or refute the team's assumptions.
• Operational Automation: Victor can perform tasks like adjusting Meta Ads budgets or processing customer refunds directly within Slack threads.

5. Notable Quotes

• "Victor is not a tool. It’s a hire." — Frederick, emphasizing that users should treat the AI with the same security and access protocols as a human employee.
• "It is unworthy of excellent men to lose hours like slaves in the labor of calculation. Let us leave that to machines." — Gottfried Wilhelm Leibniz (quoted by Frederick to frame the current AI revolution).

6. Synthesis and Conclusion

The core philosophy behind Victor is that AI agents should move away from isolated web interfaces and become integrated, proactive members of the workforce. By living in Slack, utilizing shared integrations, and maintaining a distinct personality, Victor aims to automate cognitive tasks across an entire organization. The transition from "tool" to "employee" requires careful management of security, context, and user expectations, but offers the potential to offload significant manual labor, allowing human teams to focus on higher-level strategy.