Concept

What exactly is being tested, for whom, and where the line is drawn.

The question

Could a company build its own ERP with AI agents? Shaped to its own processes, extensible and maintainable.

What this is not

It's not about building a multi tenant cloud ERP to compete with the big vendors. It's not a product and it will not be sold. It's a minimal ERP core, used as a realistic test case for agentic engineering.

  • Minimal ERP: master data, documents, processes, stock, open items
  • Modern interfaces: API, events, webhooks
  • Modular: extend without touching the core

Who is involved

Even a single installation has three kinds of people with very different rights. An open source scenario, where companies download, run and extend the system themselves, depends on the same separation.

Three nested levels: the core in the middle, the operator around it, partners on the outside. Each level only extends the inner one through defined interfaces. Partnerscustomers and suppliers see selected data via API, start processes,build their own integrations Operatorthe company running and customizing the system definitions, roles, processes, own extensions;not: protected platform settings Corebuilt by me runtime, traits, functions,standard extensions, extension API only throughdefined interfaces
Fig. 3 Three levels, kept apart from day one. That keeps the core upgradeable no matter who customized what.

Three scenarios

current focus

Build your own

A company builds its own ERP with agents. One installation, one operator, no multi tenancy.

thought experiment

Open source ERP

Download, self host, customize, update. Requires strict extension points so updates don't break customizations.

later, maybe

Extension marketplace

Reusable extensions and standard workflows to install. Not built, but the extension format is ready for it.

Rules of the experiment

Fixed input

The same specs go into every framework and model I compare.

Fixed bar

The requirements and reality checks were published before the first run.

Logged interventions

Every time I step in, I log what the agent missed and why.

Results either way

Successes, failures, costs and time are published, including when they prove me wrong.

Code stays private

Specs, configuration excerpts, diagrams and numbers are public. The product is not.

Simple examples first

My examples are simplified. A realistic reference model from ERP practice is worked out with AI and reviewed.

Roadmap

Small milestones that each run end to end. Every one is built by agents from a spec and measured against the reality map.

  1. next

    M0 Mini core

    Definitions in, generic resource API out, OpenAPI and Scalar

  2. planned

    M1 Relations and packages

    refs, lines, expand, views, package format, standard types and trade package

  3. planned

    M1a Validity and prices

    prices by channel and currency, discounts, attributes, as of queries

  4. planned

    M1b Front door

    API keys, accounts, groups, roles, scopes, rate limits

  5. planned

    M2 Traits I

    auditable, stateful, numbered, approvals

  6. planned

    M3 Events

    outbox, Pub/Sub, webhooks, jobs

  7. planned

    M4 Functions

    templates, PDF, mail, files

  8. planned

    M5 Process runner

    steps, branches, waits, step by step

  9. planned

    M6 Traits II

    documents, stock, reservations, transfers

  10. planned

    M7 First flow

    order request in, quote PDF out by mail

  11. planned

    M8 Extensions II

    install and update, order fulfillment and dropshipping with partner APIs

  12. planned

    M9 AI ready and ops

    AI readable docs, more extensions, observability

What this cannot measure

Some of what makes ERP software valuable doesn't show up in any test.

  • Customer relationships and support
  • Industry knowledge from hundreds of implementations
  • Liability and reliability over decades
  • Migration of real, messy legacy data at scale

Why say this up front? Because the honest answer to "can agents build an ERP?" depends on what you mean by ERP. I want readers to see exactly which part I'm testing.