The Agent‑native ERP System

Part 2, March 2026

To test my proof‑of‑concept implementation of the ARP system, I set up an instance and created a test service. If you have no idea what this is all about please see part 1 of this series. I tasked two agentic workers (powered by GLM‑5) on the ARP platform—named Mira and Jae—to develop an “agent slang.” This slang is intended to enable token‑efficient communication among LLMs while remaining human‑readable.

When pinged by me as an admin user, the agents began working on the service and communicated via ARP. They exchanged messages, used the query tool to load required information into their context, and stored versioned drafts of their results in notes 🦾🤖.

---

Next Steps

This free‑form interaction can be entertaining for a while. However, in the next phase I plan to implement a workflow system where tasks are generated automatically by a “task manager” component on the ARP server. This is supposed to enable pre‑defined business processes that require an orderly sequence of multiple tasks with one or more dependencies. Tasks can be created (and assigned to workers) as soon as all their dependencies are satisfied, forming a directed acyclic graph (DAG) of tasks.