Most AI orchestration tools are built around single-agent workflows. Paperclip takes a different approach: it models the entire concept of a "company" — a structured group of agents with defined roles, budgets, and health monitoring — as a deployable configuration. The result is an open-source orchestration layer that makes multi-agent coordination something you set up once rather than code from scratch each time.
What "Agent Companies" Actually Means
The central abstraction in Paperclip is the agent company — a named collection of AI agents that share a common objective. Each agent in the company has a role, a set of skills it can invoke, and a budget ceiling that prevents runaway API spending. The configuration lives in a file, which means teams can version-control their entire agent setup, fork it across projects, and collaborate on changes without needing to touch any infrastructure. Pre-built company templates let you start from a working baseline rather than blank config.
Heartbeats and Health Monitoring
One of the practical problems with running persistent agents is knowing when they've silently failed. Paperclip addresses this with a heartbeat system: agents emit regular status signals, and the orchestrator tracks whether they're responsive. When a heartbeat is missed, the system can restart the agent, alert a human, or hand the task to a backup agent. This is the kind of operational detail that's easy to ignore during development and expensive to retrofit after production.
The Skills Marketplace at skills.sh
Paperclip ships with integration hooks for skills.sh, an emerging marketplace for pre-built agent capabilities. Rather than writing every agent skill from scratch, developers can pull skills written by others — web search, data extraction, document generation — and attach them to any agent in the company. This functions similarly to a package registry but for agent behaviors rather than code libraries, and the ecosystem of available skills is expanding as more teams publish their own.
Using Claude Code as the Configuration Interface
The recommended setup workflow uses Claude Code to configure and deploy Paperclip companies. You describe the company you want — agents, roles, skill assignments, budget rules — in plain language, and Claude Code translates that into a working configuration. The feedback loop is tight enough that adjusting an agent company's structure feels more like editing a document than deploying infrastructure.
Takeaway
Paperclip occupies a useful middle ground between ad hoc scripting and heavyweight enterprise orchestration frameworks. For developers who need multi-agent coordination but don't want to build a custom orchestration layer, the model of version-controlled agent companies with budget controls and health monitoring is a practical starting point. The skills marketplace adds a path to faster iteration as the ecosystem grows. As open-source contributors add more pre-built company templates for common use cases, the time-to-deploy for new agent workflows will continue to drop.