The problem with running 80+ skills and 50+ agents from a single Claude Code session is that nothing has a paper trail. I dispatch work, it happens, and the only record is whatever I remember to log afterward. No audit trail, no budget tracking, no way to see which agent ran what, when, or how much it cost.
This week’s experiment: Paperclip, an open-source agent orchestration framework that treats AI agents like employees in a company. You create a company, hire agents into roles, and assign them work through issues. Each issue gets a full transcript, and every token gets tracked.
What I actually did: cloned it, built it, and deployed it as a persistent background service on a headless Mac Mini, exposed the dashboard through a private network for remote management, and created a single company — “JC Ops” — with four agents:
- Director (Opus) — the expensive thinker, for decisions that need reasoning
- Worker (Sonnet) — the mid-tier workhorse
- Gemini Worker — free tier, for tasks that don’t need Anthropic’s models
- Local Worker (Ollama) — running on the same machine, zero API cost
I pointed the Director agent at the vault and asked it to report what it found. It discovered the full skill library, the MCP server configs, all the instructions — everything synced over from the primary machine via Syncthing. The agent inherited the entire system without any manual setup.
Here’s where I almost made it harder than it needed to be. My first instinct was to create six separate companies — one per domain of work. That’s the kind of premature architecture that feels productive but creates coordination problems before you have coordination needs. One company, four agents, work routed through issues.
The decision that matters: all task-like work now goes through Paperclip instead of being dispatched ad hoc from Claude Code sessions. Claude Code becomes the manager layer — it creates issues, assigns agents, and checks results while the agents do the work and leave transcripts behind.
About 180 minutes from “what is this?” to a running system. Also migrated 5,107 files from legacy cloud storage into the vault during the same session, which had nothing to do with Paperclip but apparently that’s how Saturdays go now.
The transferable thing: the gap between “I have agents that can do work” and “I have agents whose work I can verify” is organizational, not technical. What was missing was structure — roles, assignments, and receipts.
Still not sure about: heartbeat scheduling. Paperclip supports agents that wake themselves on a timer, but I already have scheduled tasks running through other systems. Enabling heartbeats without auditing the overlap would mean two systems doing the same work. That evaluation hasn’t happened yet.