Paperclip gives one person a control plane for organizing AI agents into an org chart, assigning work, and setting the standards they should follow.
2
Reviewers, approvers, routines, skills, and budgets help agents complete work across coding, marketing, sales, and finance operations.
3
Dotta recommends building an organization agent by agent, using expensive models where needed and cheaper models for work that can tolerate them.
Summary
Dotta Bippa introduces Paperclip, an open-source agent orchestrator for people who want to manage AI labor through an organization rather than isolated prompts. Paperclip lets a human create an org chart, assign goals, hire agents, define instructions, and review the resulting work. Agents can use different providers and tools, while their conversations, memory, skills, and project context stay within the organization. Dotta demonstrates the system by asking a CEO agent to hire a video writer and create a Remotion video about Paperclip's GitHub stars. He also shows workflows with QA and approval stages, scheduled routines for releases, and reusable skills. His advice is to start with only the agents a business needs, improve their instructions after mistakes, and reserve stronger models for work that needs them. The early roadmap includes CEO chat, artifacts, deployments, multiple human users, sandboxing, a desktop app, and cloud deployment.
Paperclip puts a human in charge of an AI organization
Dotta describes Paperclip as a human control plane for AI labor. A person creates an org chart, sets goals, gives instructions, and stays involved from high-level design through execution. The system can have a CEO, CTO, engineers, marketing staff, and other roles. Agents can be brought in from different tools and providers, including Gemini, Pi, Hermes, OpenClaw, Claude Code, and Codex. Paperclip stores their communication and memory with the rest of the organization, so work can build on shared context instead of starting from a blank prompt.
The CEO agent can break work down through an organization
Dotta demonstrates Paperclip using the company he runs for Paperclip itself. He creates an issue for the CEO and asks it to hire a video writer and produce a Remotion video celebrating 40,000 GitHub stars. The CEO is expected to break the request down to its executive branch and individual contributors. In the demonstration, it hires the video writer, provides a Remotion skill, and uses existing project information, a dashboard, and a branding guide. Dotta says a task that could have taken him a week became an afterthought.
Skills become more useful when they include an organization's preferences
Paperclip has a built-in skills manager that can install skills such as Remotion best practices. Dotta says the agent can combine a generic skill with a Paperclip-specific one containing branding rules, preferred pacing, and style choices. He reviews the generated plan and tells the agent that cuts should be two seconds rather than six. After repeated work, the organization can collect feedback and improve its skills. In the version shown, this learning is still manual, although Paperclip is intended to automate more of it.
Review and approval workflows give agents a required path to completion
Dotta says agents often ignore informal requests such as testing a change in a browser. Paperclip can instead require a reviewer or approver after an assignee finishes. A QA agent with an agent browser skill can open a website, fill out a form, and click a button. The coder and reviewer can iterate, while a manager approves whether the result fits the organization's standards and brand. This workflow is intended to work across different agent vendors, rather than depending on hooks that vary between tools such as Claude and Codex.
Routines turn repeated operational work into scheduled tasks
Paperclip routines capture work that would otherwise live in a prompt folder or repeated copy-and-paste. Dotta shows routines that create a Discord message for items merged into the master branch, write a release changelog, or operate on a specified pull request branch. They can run on a schedule or manually, and template variables let the person provide details at run time. Routines can also use skills. In Paperclip's open-source project, one routine uses Greptile for a first-pass review of community pull requests.
Paperclip is intended for business operations beyond coding
Dotta is explicit that Paperclip is not a coding or code review tool. A coding team can use Cursor or GitHub alongside it, but a nontechnical person can use Paperclip to manage marketing, sales leads, or finance operations. He shows a report produced from saved bookmarks and describes a future version where reports could expose actions such as creating an issue, making a plan, or turning an idea into an app feature. The system is meant to coordinate work across a company.
A small organization with well-defined roles is better than a giant template
When creating a new agent-tools directory company, Dotta starts with a CEO and asks it to hire an engineer and create a hiring plan. He advises new users to add agents one at a time instead of importing a template with 130 agents and a large marketing team. Each role needs clear instructions about how it should behave, and the organization should expand only when the work requires it. He also recommends using stronger models for high-intelligence work and cheaper models through OpenRouter for tasks where they are sufficient.
Paperclip's early roadmap extends control to teams and deployments
Dotta presents Paperclip as an early release with missing pieces. The roadmap includes artifacts and deployments, CEO chat, a maximizer mode for extended agent work, multiple human users, cloud deployment, agent sandboxing through services such as E2B, and a free open-source desktop app. He also mentions work on stability, memory, and a knowledge base. At the time of the talk, multiple human users and several deployment features were still being built.
