The Startup Ideas Podcast
Episode: I Built an AI Agent Company (From Scratch)
Host: Greg Isenberg
Guest: Dota (Creator of Paperclip, via AI avatar)
Date: March 26, 2026
Overview
This episode dives into the wildly viral open-source project Paperclip—an agent orchestration layer for "zero human companies." Host Greg Isenberg speaks with Dota, Paperclip’s pseudonymous creator (appearing via AI avatar), exploring what it practically takes to build, manage, and scale a business that primarily comprises AI agents. Dota delivers a hands-on guide to getting started with Paperclip, discusses the realities versus the hype around zero-human companies, and offers insight into emerging best practices for agentic startups.
Key Discussion Points & Insights
1. What is Paperclip and Who is Dota?
- Paperclip’s Tagline: "Agent orchestrator for zero human companies."
- (01:02) “It’s a tool that you can use to put in your ideas and manage a company of agents.”
- Open-source nature and viral growth (30,000 GitHub stars in three weeks).
- Bring-Your-Own-Bot: Paperclip is model-agnostic and lets users bring their preferred agents.
- (03:36) "You can use any of these models with Paperclip. You’re going to get the best results if you use a frontier model for your CEO…"
Dota’s Persona:
- Went anonymous through NFT work; values pseudonymity.
- (02:02) "The Dota Persona has just been my online persona ever since. You know, Naval always says the best outcome is to be famous but nobody knows you."
2. Demo: Spinning Up Your First AI Company
Step-by-step Walkthrough:
- Selecting a startup idea: “Finance app that builds money habits in three minutes a day” (“Moolah”).
- Agent Assignment & Task Delegation
- Assign a CEO agent using a robust model (e.g., Claude Opus, Codex).
- Multi-agent orchestration: Each agent handles one task at a time, with a default approval flow for hiring.
- (07:51) “You are the CEO. You set direction, hire founding engineer, write a hiring plan, break the roadmap into tasks.”
Spend, Tracking, & Accountability
- Transparent token budgeting and spend tracking built-in.
- (09:31) “It actually tracks your monthly spend...I built Paperclip because I was using cloud code...I couldn't remember what any of them were working on...My budget was busted and I didn’t know what was actually accomplished.”
Mindset Shift
- Shift focus to high-level board/CEO thinking; set goals and let agents implement.
- (11:23) “...the goal of Paperclip is you’re thinking at a higher level. You’re the board… Here’s the MRR we want to reach, here’s what we want to achieve… Now make a plan.”
3. Configuration & Agent Design Patterns
Persona and Memory:
- Agents use a memory system (e.g., PARA memory); users must give clear, persistent context ("Memento Man" analogy).
- (13:05) “Your AI agents are Memento man. They wake up...they don’t know who they are, what they’re supposed to be doing. So you need to ‘write it on their arm’ with good instructions.”
Prompting, Instructions & Skills
- Key to high-quality agent work: explicit instructions, context, and clear success criteria.
- (14:41) “If it's forgetting to set a win condition… I'll literally just add that line...and you'll get better results.”
- Installing skills to agents (e.g., Remotion for video editing).
- (19:17) On third-party skills: “They have badges...that’s as best as you can do probably for now.”
Agent Safety & Skill Security
- No foolproof solution for risky third-party skills—look for community trust and security badges.
4. Evolving Company Structures & Tasks
-
Scaling up: Add video editors, QA, marketers by simply prompting the CEO agent.
-
“Firing” agents and managing org structure possible in the interface.
- (25:33) “Right here we got terminate. You can terminate.”
-
Discussion of agentic org patterns (“harness engineering”)
- Ensuring checks and balances (e.g., “engineer asks QA for review”).
- Structure is still evolving; early best practices forming.
-
Taste & Communication:
- Instilling organizational “values” and “taste” into agents is the new frontier for effective leadership.
- (29:42) “AI can do everything except know your values. You have to become more aware of your values and find out how to communicate them back.”
5. Paperclip in Practice: Real-World Usage
Paperclip’s Own Paperclip Org Chart:
- CEO with direct reports (CMO, CTO, UX, QA engineers, etc.).
- Uses different models for different agent types (e.g., Claude, Cursor).
- Adding new org features: evaluation, learning from feedback, routines (like daily Discord auto-posts).
- (33:27) “We want to have a Discord channel where we post the updates … especially want to celebrate our community members who contribute.”
Examples in the Wild
-
Not just tech startups: Security review firms, dentists’ foundations, roofing companies using Paperclip for automating real-world work and managing leads.
- (37:37) “Even like blue collar… we’re exploring using Paperclip to help their sales agents find leads…”
-
Import & export company templates:
- E.g., Gary Tan’s “G Stack,” Superpowers repo, agency_agents—import rich org structures & skills to jumpstart businesses.
- (39:12) “You can import them all into Paperclip with all the same skills… Actually referencing these remote repos…”
6. The Future & Limitations of Zero-Human Companies
- Reality Check: Completely automated companies are "aspirational," not yet the default—agents still need human high-level guidance.
- The ‘Maximizer’ Mode: Upcoming feature for relentless agent execution until goals are met (token spend permitting).
- Scale & Survivability:
- Paperclip is designed to evolve even as models get more advanced—agents still need organizational context, structure, and human taste.
- (44:34) “I wanted it to do well because I knew it was good… it struck a chord with people… Will you need something like Paperclip in a year? … I think you still need software to help manage the jobs that are left.”
Notable Quotes & Memorable Moments
-
On the Power of Context for AI Agents
- Dota (13:11):
“Your AI agents are Memento man… they don’t know who they are or what they’re doing. So you need to give them little Polaroids or write tattoos on their arm.”
- Dota (13:11):
-
On Why Paperclip Went Viral
- Dota (44:34):
“It really struck a chord with people. The big question is: will you need something like Paperclip in a year from now? ... I do think Paperclip will have a bit of staying power because it’s not fixed to one agent. It’s operating at this higher level of ‘what do you want to do?’”
- Dota (44:34):
-
On Instilling Taste & Leadership
- Greg (29:42):
“Even in a pre-AI era, the concept of a good leader… is someone who can clearly communicate their values and taste.” - Dota (29:59):
“That’s right.”
- Greg (29:42):
-
On Community and Openness
- Dota (43:26):
“There’s myself, Dota, and two cofounders, Devin and Scott, and then also the community, who is just doing an incredible job at contributing every day.”
- Dota (43:26):
Timestamps for Key Segments
- 00:00–04:00 — Project intro, Dota’s anonymity, what Paperclip is
- 07:00–15:00 — Demo: starting a company, agent setup, memory, practical advice
- 18:49–21:30 — Skills, agent security, value of GitHub stars vs. trust
- 25:17–30:00 — Company structure, firing agents, communication of taste
- 30:37–36:30 — Inside Paperclip’s org chart, real routines, Discord updates
- 36:35–41:00 — Real-world usage and niche applications
- 41:56–44:00 — Sharing org templates, importing “proven teams”
- 44:05–46:00 — The future, maximizer mode, Paperclip’s staying power
Conclusion
In this hands-on, honest episode, Dota makes it clear that the “zero human company” dream is closer than we think, but requires new mental models—think “AI as employees,” with humans stepping into more abstracted, high-level leadership roles. While agent technology is evolving fast (and Paperclip leads that charge), successful deployment still depends on thoughtful setup, persistent context, robust evaluation, and most importantly, clear communication of human values.
Try Paperclip: Paperclip on GitHub
Startup idea database referenced: gregisenberg.com/30startupideas
