I’m testing a smaller version of this around a live website, although the agents don’t have autonomous production access.

The first thing that broke was goal state: a new KPI arrived while agents were still working from the old one, so their output looked valid but was already stale. That pushed me toward narrower roles, explicit human authority and blocking material changes until they’re approved.

I documented the setup and its current limits here: https://github.com/nohuman-labs/agent-company

I’ve been using Hermes to test a small “agent company” around a real website: a human sets the KPIs, while separate CEO, Product, Growth and Engineering profiles coordinate through a shared Kanban board.

It’s still an experiment, and I’ve documented what it can and can’t do here: https://github.com/nohuman-labs/agent-company