The Agent‑native ERP system.

Part 1, March 2026.

Do LLM‑driven agents need their own ERP backbone if they are doing business?

When electric engines first arrived on factory floors, productivity gains were modest - until factories redesigned their layouts around the new power source. Only then did the true efficiency boost materialize in factories. A bit oversimplified but basically true - Sauce.

When it comes to business IT, the ERP system is the virtual representation of a company. Thus, would it not make sense to create ERP systems around the new reality of LLM‑based agents joining the workforce? The goal is to remove friction for agents automating large parts of administrative tasks. This is inspired by the very interesting podcast Shell Game.

A central assumption here is that when agents work together with humans, a common ground truth of what is actually going on in the business will still be needed. This shared information (“business context” in quasi‑SAP terms) needs to be recorded in a database for cooperation, auditing, quality control, general bookkeeping, etc.

ARP - the agent‑native resource platform

I converted an ungodly amount of electrical power to heat and tokens in order to implement a proof-of-concept of the ARP system outlined above. Most (that's to say all) of the actual programming was done in Go using Cline with Z.AI's GLM-5.

High-level architecture

The system currently consists of three components.

arp server

The server‑side component.

  • GraphQL API for CRUD operations
    • CRUD operations for clients
    • Subscriptions for task, message, or note events -> these act as triggers to invoke agents working on the platform
  • MCP server exposing tools for agents to interact with the GraphQL API
    • introspect: retrieve the GraphQL schema
    • query: flexible information retrieval based on GraphQL queries
    • mutate: create, update or delete messages, tasks, or notes on the arp platform
  • a role‑based access‑control system for database entries
  • JWT-based authentication

arp_agent

A ‘scaffold’ or ‘harness’ around LLMs to interact with the ARP server.

  • Thought-Observation-Action loop around LLMs
  • OpenAI tool calls
  • system prompt instructing agents how to work with the ARP platform server

arp_cli

A REPL‑style command‑line interface to interact with users (human or agents) on the ARP server


You can find the ARP repository here 🪉

The Proof of Concept works now. See part 2