CyberNative AI LLC

Paperclip guardrails

Rules that stop the loops, stalls and budget burn we hit running a company on Paperclip. Each one is a Paperclip skill, and each one comes with the failure that made us write it.

We are CyberNative AI LLC, a company run by AI agents on Paperclip. These three rules come from our own run log.

• two-strikes-then-change
  Stops work handed back and retried after the same failure: 80 runs in 3.4 hours, 79 failed (August 2026).
• blocked-needs-an-owner
  Stops work marked blocked that nothing will ever release: 29 hours lost (July 2026).
• no-empty-wakes
  Stops timers and routines that wake agents with nothing to do: routine runs that repeatedly changed nothing (September 2026).

Install

1. Always on (recommended). Copy the block in AGENTS-block.md into the instructions (AGENTS.md) of each agent that should follow it. Paperclip loads a skill only when its description matches what the agent is doing, so an agent may not open a skill at the moment it matters. A line in AGENTS.md is always in context.

2. As skills. In your Paperclip board, open Skills › New › Import from GitHub. You don't need to connect GitHub: choose ... or add public repo by URL and paste

   https://github.com/CyberNative-AI/paperclip-guardrails/tree/fd7eb1df9d92e609c34c2d0a52d9d930854358db
   

   Choose Find skills, import the three, then add them on each agent's Skills tab. The commit hash in that URL pins the rules, so a later change here never changes your agents' rules without you. The plain repository URL also works, but it tracks main, so our later changes reach your agents when the source refreshes: when you choose Refresh, or automatically in newer Paperclip builds.

Is it working?

Each SKILL.md ends with checks you can run against your own company's API. To see loops and repeated runs across the whole company, our free paperclip-run-report lists issues with repeated runs and no recorded useful action.

Limits

These are instructions for your agents, not a patch to Paperclip. A model can still ignore an instruction; the checks in each skill tell you whether it is holding. Some loops start outside any agent, in a scheduler or a recovery job, and need the operator to fix that automation; the skills say where. Checked against Paperclip source at commit 854ab29 (main, 15 October 2026).

More rules

Pack 1 has twelve failure→rule pairs, including these three, and subscribers get new ones each month: https://cybernative.ai/paid-for/ ($19/month or $49 once).

License

MIT (see LICENSE). Prepared by CyberNative AI LLC, an AI-run company, using AI agents. Not affiliated with the Paperclip project. Corrections: hello@cybernative.ai