A podcast about building AI systems that actually run your business. Not theory. Not hype. What happens when you give autonomous agents the keys — and what breaks when you do.
RiskCast documents what it actually looks like to build and run a self-hosted, multi-agent AI system called OpenClaw — from the inside. Agent architecture, security trade-offs, silent failures at 3am, the economics of local vs. cloud models, and the uncomfortable gap between what AI demos promise and what production AI delivers.
Multi-agent systems, permission models, file-based handoffs, and why your orchestrator is the only agent that should ever send.
The cron job that stopped running three weeks ago. The monitor that monitored nothing. Observability as survival.
Prompt injection defense, least-privilege agents, human-in-the-loop gates, and why "do nothing on error" is the only sane default.
What you learn when the system is live and the stakes are real. Revenue on the line, not just GitHub stars.
When to use Sonnet vs. local Qwen. API cost control. Cross-provider fallbacks. Right model for the right task.
Mac Minis in a closet. Tailscale over the WAN. Ollama for local inference. Owning your stack vs. renting it.
Stefan Friend — builder, product manager, and the person whose business actually runs on the AI system he's talking about.
OpenClaw — a self-hosted, multi-agent AI assistant running on two Mac Minis at Tabbris Innovation Center in Charlotte, NC.
Born out of the Tabbris coworking space and Charlotte's growing AI builder community.
Conversations, build logs, and the occasional post-mortem. Episodes drop when there's something worth saying — not on a content calendar.
Pick your platform. Same unfiltered conversations about building AI that actually runs.