The Startup Ideas Podcast

I Built an AI Agent Company (From Scratch)

254 snips
Mar 26, 2026
An AI company gets built live, complete with a CEO, engineer, QA, video editor, and content strategist. The conversation explores agent memory, heartbeat checklists, skills, and QA loops. It also gets into token budgets, recurring routines, shareable AI companies, and the big question of whether agent orgs actually work.
Ask episode
AI Snips
Chapters
Transcript
Episode notes
INSIGHT

AI Agents Need Memento Style Memory Scaffolding

  • AI agents are highly capable but identity-less, so they need persona prompts, memory, and a startup checklist every time they wake up.
  • Dotta calls them Memento Man and uses heartbeat steps like fetch who you are, read today's plan, and extract memory.
ADVICE

Patch Agent Behavior By Updating Rules Fast

  • When an agent repeats a mistake, edit its operating rules directly instead of hoping the next prompt fixes it.
  • Dotta adds lines like define success conditions and require QA review, then asks the CEO to hire missing roles.
ADVICE

Use Skills To Extend Agents But Audit Them

  • Install reusable skills to expand what agents can do, but treat public skill repos as partially trusted, not safe by default.
  • Dotta points to skills.sh badges and GitHub stars as directional signals, then installs a Remotion skill into the company.
Get the Snipd Podcast app to discover more snips from this episode
Get the app