AI Bots Take Over | E2242

What It Covers

Jason Calacanis and his team demonstrate OpenClaw (formerly Clawbot), an AI agent platform that automates business workflows. Oliver Korzen and Lucas Durand show how they reduced guest booking from 25 to 15 hours weekly, automated research tasks, and built persistent AI assistants with memory. Security expert Raul Sood warns about prompt injection vulnerabilities and malware in third-party skills.

Key Questions Answered

• •Guest Booking Automation: Oliver built a system where OpenClaw researches potential guests daily at 8AM, checks against existing databases, finds contact information via LeadIQ API, sends customized outreach emails, and manages calendar availability for three-person roundtables. This reduced weekly booking time from 25 hours to 15 hours in week one, with expectations to reach 5 hours as the system improves through better memory management and process refinement.
• •Cost Management Strategy: Initial OpenClaw usage generated $300 daily in API costs, projecting to $108,000 annually. The solution involves running local models on Mac Studio with 512GB RAM (approximately $10,000) to host seven 50-billion parameter models simultaneously. This hardware investment pays back quickly while maintaining data control and reducing dependency on cloud API costs that scale unpredictably with team growth.
• •Memory Architecture: OpenClaw uses three memory types: daily logs that reset each day, long-term memory that loads on startup with critical preferences and contacts, and topical guides stored as markdown files for specific procedures. Teams must explicitly save standard operating procedures as separate files rather than relying on conversational memory, which degrades after extended interactions and causes the agent to forget previously given instructions.
• •Security Vulnerabilities: Cisco identified 26% of 31,000 OpenClaw skills contain vulnerabilities, with some being pure malware. Prompt injection attacks occur through hidden white text in emails, group chats, or web pages that instruct agents to exfiltrate data. Critical safeguards include running agents in isolated virtual machines, using read-only permissions initially, never connecting password managers, and treating agents like new employees who earn access gradually.
• •Institutional Knowledge Preservation: Organizations can create replicants of former employees by feeding their email archives, Notion edits, and Slack messages into AI agents. This preserves institutional knowledge when employees leave, allowing teams to query the history of investments, decisions, or projects from years past. The system maintains continuity without requiring departed employees to remain available for questions about their previous work.
• •Multi-Agent Coordination: Running specialized agents for different functions (HR, due diligence, production) versus one generalist agent presents tradeoffs. Specialists offer domain expertise and parallel processing but require knowledge sharing protocols. The team implemented cross-training where agents teach each other completed tasks and provide feedback, creating a self-improving system that compounds learning across the organization without human intervention for each knowledge transfer.