Getting Started

Getting Started

You need exactly three things to stop the worst multi-agent failures. This takes about 10 minutes.

Run the reference first

git clone https://github.com/ElmatadorZ/genesis-governance-os.git
cd genesis-governance-os/reference/python

python demo.py              # the 3-agent workflow, executing live
python conformance_test.py  # 18 invariant checks — proves the spec holds

No dependencies, no API key, no model. Python 3.9+ only.

demo.py runs Research → Write → Review and prints the audit trail — every permission verdict, every handoff with its note, and the close. That trail is the whole point: you can see what happened and why.

Then implement three standards in your stack

This is a standard, not a framework — you implement it in whatever language you already use.

Order	Standard	Implement
1	GTS-1 Task State	six fields, single owner, append-only attempts, explicit done_when
2	GPS-2 Permission	deny-by-default, gate before the tool call, irreversible ⇒ human
3	GOP-3 Orchestration	route by fit, hard loop guard, human-visible close

That is a conformant system. A 3-agent line with done_when on each task and a loop guard is enough to start.

The minimum that actually prevents incidents

If you only do two things today:

1. done_when on every task — an agent cannot declare victory it has not earned.
2. Deny-by-default with a pre-execution gate — the agent cannot send the email, run the destructive command, or publish, unless you explicitly allowed it.

Those two alone remove most of the failure modes that make multi-agent systems unusable in production.

When to add governance

Add GOS-0 Constitution when the number of agents grows past what one person can watch. Before that point it is overhead; after it, its absence is how a system quietly becomes something you did not intend.

Then claim it

When you are ready to state compliance publicly, follow Conformance — declare your level (L1/L2/L3), publish your suite result, and attest the requirements a test cannot prove.

---

📖 QUICKSTART.md — a worked 3-agent example