I Built an AI Agent Company (From Scratch)

What It Covers

DOTA, creator of Paperclip — an open-source AI agent orchestration platform that reached 30,000 GitHub stars in three weeks — demonstrates how to structure, configure, and manage a multi-agent company using the tool, covering agent hiring, memory systems, skills, routines, and cost tracking.

Key Questions Answered

• •Agent Memory via Heartbeat Checklists: AI agents have no persistent identity between sessions — treat them like the protagonist in Memento. Configure a heartbeat checklist in each agent's persona settings that instructs them to confirm their role, read today's plan, locate assignments, and extract memory before beginning any work session to maintain continuity and output quality.
• •Frontier Models for CEO, Cheaper Models for Tasks: Assign a frontier model like Claude Opus or Codex to the CEO agent role for high-level planning and delegation. For subordinate tasks — content drafting, QA checks, routine data gathering — use free or low-cost models available on OpenRouter, where models like Hunter Alpha and Step Flash have periodically been available at zero cost.
• •Iterative Prompt Refinement via Persona Rules: When an agent repeatedly produces substandard output, add a single corrective rule directly to its persona configuration rather than re-prompting each time. For example, adding "define a success condition for every task and request QA review before closing" to an engineer's persona eliminates recurring quality gaps without manual intervention on each issue.
• •Routines for Recurring Agent Tasks: Paperclip's routines feature converts repeating workflows into scheduled issue templates with full token-spend tracing. Set a daily trigger — such as 10AM — to have a content agent read the last 24 hours of GitHub commits, draft a community-formatted Discord update, and post results directly in the issue log for review before publishing.
• •Import Proven Agent Organizations via GitHub Repos: Rather than building agent teams from scratch, import pre-structured company templates from public GitHub repositories directly into Paperclip. Repos like the 60,000-star Awesome Agents collection contain 100-plus pre-configured agents with skills attached. Paperclip references the remote repo, so imported agent teams receive upstream updates automatically without manual reinstallation.