Paperclip: Hire AI Agents Like Employees (Live Demo)
TL;DR
Paperclip frames AI agents as an org chart, not a chat box — Dota describes it as an “agent orchestrator for zero human companies,” where you define goals, hire a CEO and specialists, track budgets, and manage work through issues instead of juggling 20–30 Claude Code windows.
The core trick is giving agents memory and a heartbeat — Dota compares AI agents to the protagonist in Memento: capable but disoriented, so Paperclip uses persona prompts, file-based memory, and a startup checklist to remind each agent who they are, what matters, and what to do next.
Best results come from frontier models at the top and cheaper models below — He recommends Claude Code or Codex for the CEO role, then lower-cost or even free OpenRouter models like Hunter Alpha or Step Flash for narrower tasks, with spend tracked at the company level.
Paperclip’s real value is visibility and control, not full autonomy yet — The demo shows manual approval for hiring, token-spend tracking, issue histories, routines, skills, and QA handoffs, all meant to avoid the classic agent failure mode where things “fall apart” after the first 30 minutes.
The strongest early users aren’t building sci-fi startups — they’re augmenting existing businesses — Dota says in Paperclip’s first three weeks, users have applied it to a security review firm, a dentist’s foundation and family operations, and even a roofing company hunting hail-damaged leads with satellite and weather data.
Paperclip’s big bet is that orchestration survives model progress — Even if raw models get dramatically better, Dota argues people will still need software to manage taste, roles, feedback loops, and organizations of 10 or 100 AI workers, which is why Paperclip is model-agnostic and “bring your own bot.”
The Breakdown
From viral GitHub repo to “zero human company” pitch
Greg opens by framing Paperclip as one of the breakout open-source AI projects of the moment: 30,000 GitHub stars in three weeks. Dota, appearing as his longtime avatar persona, defines it simply as an agent orchestrator for “zero human companies” — a system for assigning roles, goals, and work to AI employees that run around the clock.
The live demo starts with a startup idea and a CEO hire
They pick a sample idea from Greg’s browser — a finance app that builds money habits in three minutes a day — and name it Moola. Dota shows the initial workflow: define the company goal, then hire a first agent, usually a CEO running locally via Claude Code or Codex, because right now Paperclip works best on your own machine.
Why Paperclip exists: too many agent windows, no accountability
Dota explains he built Paperclip after trying to run companies with cloud coding agents and ending up with 20 or 30 windows open, no clue what each one had done, and blown token budgets by Monday. Paperclip is his fix: track monthly spend, route work through issues, and ensure one task belongs to one agent at a time so they don’t collide.
The Memento analogy for agent memory is the real unlock
When Greg asks how to configure agents, Dota lands on the most memorable part of the conversation: AI agents are like the guy in Memento. They wake up highly capable but with no idea who they are or what they’re supposed to be doing, so you need “little Polaroids” and tattoos — heartbeat checklists, memory files, safety rules, and explicit instructions about success conditions and QA.
Skills, context, and why top-tier output still needs human taste
Dota shows how Paperclip can install reusable skills like Remotion best practices via skills.sh, while admitting security is still a real concern and nobody has fully solved it. When Greg asks how to get a “top 1% video editor” result, Dota is blunt: Paperclip can help organize context, but it can’t invent your brand guide, references, or taste — AI can do almost everything except know your values.
From founder micromanagement to actual org design
A small UI bug in Paperclip itself becomes a nice reality check: Dota manually files an issue about misaligned mention pills and asks engineering to pass it to QA with a screenshot. He says that’s obviously not how a true zero-human company should work, which is why the real challenge is designing organizations where design, QA, and review agents catch those things without the founder acting like a design PM.
How Paperclip runs Paperclip — and where routines come in
Dota pulls back the curtain on his own internal org: a CEO with a CMO, UX designer, CTO, QA engineer, coder agents, and an emerging eval engineer meant to run “performance reviews” on agents. He demos a routine that checks GitHub merges every day and drafts a Discord update celebrating contributors, emphasizing that Paperclip treats recurring jobs like auditable issues instead of invisible background automation.
Early traction, imported companies, and the open question about the future
The most interesting customer stories are surprisingly practical: a security review company, a dentist, and even a roofing business using agents to identify hail-hit neighborhoods with better sales odds. Dota then shows where Paperclip is headed — importable companies inspired by Gary Tan’s Gstack, agency repos with 100+ agents, and a future “maximizer mode” where the CEO keeps staffing and pushing until the work is truly done — while conceding the biggest question is still whether orchestration software like this remains necessary as base models improve.