What happens when you treat an AI agent like a co-founder instead of a tool?

In this episode, Valentino and Joe go deep into a real-world experiment: spinning up an autonomous agent using OpenClaw, giving it domains, goals, and just enough guidance to build an actual business. From creating accounts and managing projects to writing code, deploying with Kamal, and even designing its own training curriculum, the agent evolves from confused assistant to something resembling a junior engineer with initiative.

Along the way, they explore the messy reality of agent workflows: memory systems, self-training loops, PR reviews, hallucinated confidence, and the constant tension between autonomy and control. The result? A working product, 15 early users, and a pile of hard-earned lessons about what AI can and definitely cannot do today.

If you’re building with agents, thinking about autonomous systems, or just curious what happens when you let AI run a startup… this one’s for you.

🔗 Show Notes

- Valentino's Minerva Experiment
- Minerva's First Product

Core Tools & Frameworks
- RubyLLM (Carmine Paolino)
- Kamal (Deploy Rails anywhere)
- Tailscale (Secure networking)

Libraries & Infra Mentioned
- ExtraLite (SQLite performance layer)

Learning & Community
- Ruby AI Newsletter (Matt Solt)

Other Mentions
- OpenClaw
- Claude Code
- Action MCP
- Fizzy (37signals)
- Magic Beans (graph-based project management for agents)
- ups.dev (agent status pages project)
- DailyVibe.ai

Books & Resources Referenced
- Practical Object-Oriented Design in Ruby by Sandi Metz
- Programming Ruby (Pickaxe Book)
- The Well-Grounded Rubyist
- Layered Design for Ruby on Rails Applications — Vladimir Dementyev

Cultural Reference
- Wired article on AI-generated band marketing (“Geese”)

00:00 Podcast kickoff
00:40 Geese AI marketing psyop
02:03 Starting an AI band
04:18 Daily Vibe artist generator
06:24 Open Claw origin story
08:52 Domains to business ideas
10:44 Onboarding an AI coworker
13:22 Handholding and action loops
14:10 Shark Tank idea filter
15:32 Training and memory system
20:20 UPS dev agent status pages
22:54 Rails build struggles
24:04 Bootcamp with Ruby books
26:20 Rebuild MVP and open source
27:55 Deploying with EC2
28:38 Locking Down Access
30:06 AI PR Reviews
32:50 Self QA Automation
36:11 Fixing Agent Memory
38:15 Email and Token Costs
40:11 Heartbeats and Delegation
43:01 Customer Discovery Lessons
44:49 Selling Workflow Friction
48:19 Knowledge Base Frameworks
51:37 Open Source Model Future
53:57 Security Agents and Wrap

Valentino Stoll and co-host Joe Leo open the Ruby Podcast noting OpenAI is winding down its SOA video app and discuss the broader difficulty of building AI businesses. Guest Bekki Freeman, staff software engineer at Caribou Financial and organizer of Rocky Mountain Ruby, shares conference details (Boulder, Colorado at eTown, September 28–29; CFP opening soon; tickets after the schedule). The conversation focuses on safely scaling AI use in an 8-year Rails monolith: preparing messy codebases with dead code and metaprogramming, strengthening test harnesses and coverage, improving documentation, and being explicit about desired patterns rather than copying existing bad ones. They discuss PR review bottlenecks from increased AI-generated PRs, ideas like specialized AI review agents, stronger RuboCop rules, pairing/mobbing, and remote knowledge-sharing practices, plus security cautions and what AI may and may not replace (tech-debt work vs “taste”).

00:00 Sora Shutdown News
00:57 AI Hype Reality Check
01:43 Meet Bekki Freeman
02:00 Rocky Mountain Ruby Update
04:32 AI Meets Legacy Rails
07:22 Prep Codebase for AI
10:06 Patterns Versus Best Practices
12:37 Testing Strategy and TDD
16:45 PR Review Bottlenecks
19:27 Specialized Review Agents
21:31 Defining Quality Context
24:29 Humans and Team Adoption
25:20 Remote Change Adoption
27:00 Creating Sharing Rituals
29:19 Release Calls As Watercooler
30:12 Mob Sessions With Agents
33:55 Security And YOLO Risks
35:45 Too Much Code Problem
37:16 Vibe Coding Vs SaaS
42:10 AI Engineering In Two Years
45:33 Codex Versus Claude
47:39 Wrap Up And Farewell

On the Ruby AI Podcast, hosts Valentino and Joe Leo welcome Scholarly CTO/co-founder Kelly Sutton to discuss building a vertical SaaS “faculty information system” for universities. Sutton explains why competitors can’t easily replicate Scholarly: higher ed is moving off decades-old homegrown software, and the product must meet trust, security, compliance, and regulatory demands such as SOC 2 Type II. He describes how Scholarly expanded from replacing Excel/Access tracking to sophisticated workflow automation and how universities recently shifted from AI skepticism to AI FOMO. Scholarly uses AI in product surfaces, heavily in engineering, and via an admin MCP server that helps ops/customer success rapidly configure workflows from faculty handbooks with human-in-the-loop review. The conversation debates MCP’s likely temporariness versus traditional APIs, emphasizes smaller reviewable “PR-sized” outputs, and frames AI as an implementation detail focused on customer value. Valentino also shares an experiment training Claude to build products, including ups.dev and an open-source Ruby uptime-monitoring gem.