I spent seven years as a managing director in a regulated pharmaceutical
wholesale business. P&L responsibility, compliance audits, dunning letters
at 22:00 because a payment was late. I am not a developer.
Four months ago I started building an open-source multi-agent operating
system anyway. ARCHON OS is MIT-licensed, self-hosted, and it just
released v0.5.4.2. Here is why a business executive built it, what it
does, and what it honestly does not do yet.
showdev
Why AI projects fail in mid-sized companies (and it is not the models)
When AI initiatives fail in small and mid-sized businesses, the model is
rarely the problem. The business problems are:
- Uncontrollable cost. Nobody can answer "what did this task cost us?" before running it.
- Data protection. Finance, contracts and customer data do not belong in someone else's cloud.
- No auditability. An AI decision without a verifiable trail is an auditor's nightmare.
Those are not research problems. They are operating problems — which is
why a Kaufmann's answer is an operating system, not a prompt collection.
!(https://dev-to-uploads.s3.us-east-2.amazonaws.com/uploads/articles/djgratmeocr7e9kcuix4.png)
What ARCHON OS does
It orchestrates specialized agent profiles across 8 business domains
(tax/law, procurement, sales, tech, medicine, real estate, content,
strategy). Each profile has a 3-tier prompt: compact, pro or full — agents
spend tokens when their domain is actually needed, not always.
The parts I would expect readers here to care about:
A central router with cost control. Every request goes through one
switchboard (OmniRoute). Budgets are checked before dispatch, per task and
per model. If a request exceeds the budget, it is blocked — not billed.
Zero-egress, where it counts. Finance, contract and CRM work stays on
local models (Ollama, LM Studio). Worker containers run in an internal
Docker network with no internet route. The route does not exist instead of
being forbidden.
A hash-chained audit log. Every dispatch is recorded with provider,
model and host in a SHA-256 chain. Records cannot be altered unnoticed.
That is what "revisionsfähig" (audit-proof) means in German accounting law.
Kanban with a four-eyes gate. Agents propose, a human approves,
rejects with a reason, or the task escalates on expiry. Nothing goes out
unchecked.
What it looks like
Mission Control (operations dashboard):
AI Boardroom debate (55 decision frameworks, OKR to pre-mortem):

Try it without an API key
git clone https://github.com/michaelbeatman/ARCHON-OS.git
cd ARCHON-OS
docker compose up -d
Or the zero-key, read-only demo:
ARCHON_DEMO_MODE=true docker compose -f docker-compose.demo.yml up --build -d
# then open http://localhost:8501
Honest status
Pre-1.0. Interfaces change between releases, the ERP/CRM module is alpha,
and multi-user role enforcement is still being hardened. The README says
this up front and I repeat it here. If that does not scare you off, issues
and PRs are genuinely welcome — including the uncomfortable kind.
Full docs, a 97-page manual and the changelog are in the repository:
https://github.com/michaelbeatman/ARCHON-OS
A one-page overview with screenshots:
https://michaelbeatman.github.io/ARCHON-OS/
The part I did not expect
Building this as a non-developer changed how I think about AI systems.
The hard part was never the model. It was the same thing that makes a
real company run: who may decide what, who checks it, and who can prove
afterwards that it happened that way.
If you are building agents, I would love to hear how you handle the
boring half — budgets, approvals and audit trails. That is where I spend
my time.




Top comments (0)