[ erp system × ai agents ]
A full open-source ERP, exposed entirely as an MCP server and therefore completely operable by AI agents. In the pilot business, a butcher shop with a counter, vending machines, and an online shop, employees say what they need in plain language; the agents do the work in the systems. This page shows how the system was developed, the decisions behind it, and the experience gathered along the way.
Agents with system access and skills do the work inside the system.
Several mechanisms secure agent operation in the real world.
Integrated as an MCP server so agents know what to do.
Setup and interaction along the value chain, in natural language.
[ system operators of the future ]
The common expectation goes: once AI agents take over the work, business software becomes obsolete. My experience from this project is the opposite. The more an agent does on its own, the more important the system underneath becomes. It delivers exactly what a language model cannot: it enforces completeness, because an incomplete transaction simply does not happen. It ensures consistency, because bookings are interconnected. And it produces the audit trail as a byproduct.
An agent working next to the systems invents its own truth. One working inside them cannot. The roles shift: the human becomes the decision maker, the agent the operator, and the system remains the source of facts. That is why, to me, the systems with their rules and limits are themselves a governance mechanism.
[ the full value chain, end to end ]
Deploying individual AI agents in isolated areas is not the challenge; it is the reason behind the creation of new islands. The value emerges when the technology integrates along the entire value chain and the agents hand tasks to each other as specialists.
[ Sales ]
The sales agent merges the sales data of all channels and builds the daily demand forecast per article, enriched with weather and holidays. Before: collecting numbers from several systems and copying them into spreadsheets by hand.
[ Production ]
The production agent traces demand back through recipes and BOMs to raw materials, builds the daily plan, and books the confirmations. Before: planning by experience, confirmations on paper at closing time.
[ Purchasing ]
The purchasing agent creates orders: supplier, quantity, date, and terms from history and framework contracts. What it can prove from data goes through; what it can only assume goes to a human as an approval. Before: phone calls, shouts across the room, and typing it in later.
[ Warehouse ]
The warehouse agent books goods receipts, monitors stock, batches, and expiry dates, and reports coverage. Before: counting lists on paper and manual transfer.
[ Finance ]
The finance agent books the documents of all channels, monitors open items, and prepares reports, every booking with an audit trail. Before: typing up stacks and end-of-month rush.
[ HR ]
The HR agent puts the demand forecast next to the shift plan and absences: how much needs serving tomorrow, who is out, who can fill in where. Staffing emerges from numbers already in the system. Before: a pinboard and a phone chain for every sick call.
[ architecture ]
The most important architecture decision of this project: separating the levels strictly. The agents and the AI models are deliberately kept outside, because this technology evolves too fast for fixed bets. Governance sits in its own layer that no agent can bypass. And the business logic is not rebuilt, it is used.
A supervisor plans, distributes the work to specialized subagents, and merges the results. The knowledge lives in versioned skills that are loaded only when needed. That keeps the context small, saves cost, and allows any number of capabilities without maintaining every agent individually.
Why LangGraph: the process is a fixed plan with clearly defined transitions, the intermediate state is stored permanently, and interruptions for human decisions are built in. An approval can take three days. A conversation thread does not survive that; a stored state does.
Still: the framework is not the core of the solution. The logic lives outside the framework and stays replaceable when the next generation of tooling arrives.
Every agent action passes through the gateway; there is no way around it. What an agent may do is defined. What it does is logged without gaps and stays traceable down to every step. What is uncertain goes to a human as an approval card. Agent operation stays safe even when a model gets it wrong.
The tool index lives here too: the gateway discovers the entire ERP surface automatically and makes it searchable. Language models do not execute by themselves; they understand functions and choose the right one. No model manages that over a flat list of thousands of entries, so the agent works with two tools: one finds the matching function through a search by meaning and keywords, the other executes it.
This layer is independent of the agent framework and of the ERP. It is the reusable core that transfers to any further system.
The open-source ERP runs unchanged in its standard form and therefore stays updatable at any time. Over 3,400 standard functions and document types are available to the agents without changing anything in the system itself: the discovery is done by the gateway.
The business logic does not have to be rebuilt, it is already there. A transaction only happens when all information is complete. Required fields, checks, and booking rules act as a built-in guardrail for the agents, without defining workflows anywhere.
[ perspective ]
Today's language models and agent technology can already do a lot. Waiting for the next bigger or more expensive model adds no further value in my view: the capabilities are sufficient when architecture and data foundation are right.
The real bottleneck sits in the data and systems. Whoever did not prepare them cleanly for humans in the past will struggle even more preparing them for language models. The standards for this exist and are available, MCP above all. The use cases and solutions on the market build on them insufficiently so far.
How much easier it gets with these standards shows in the setup: because the agent reaches the full range of functions via MCP, it works along on structure and master data. In the internal test, a complete company stood in the system after twelve chat messages. And in daily operation, the agents support along the entire value chain, directly in the systems.
[ numbers ]
over 3,400
standard ERP functions and document types, exposed for AI agents
~95%
reliability in choosing the right tool, no hallucinations
any model
usable, chosen cost-efficiently per type of task, automatically in the system
∞ AI agents
and skills definable without managing each one: the skill carries the knowledge
[ technology ]
Wondering how this technology could create real value in your business? Then let's talk: you bring the challenge, I bring the experience and the tools.
