The model is not the product. The control plane is.
Models change, APIs move, local hardware gets upgraded, and fashionable frameworks rotate every quarter. The durable part is the operator surface: how work is requested, delegated, verified, remembered, and shipped.
What the control plane does
Hermes ties the working lanes together: Discord as the primary operator interface, local tools for files and browser verification, scheduled jobs for recurring checks, memory for durable facts, and bounded delegation to specialist workers.
That means the operator does not need to remember which script lives where every time. The control plane knows enough to route work without pretending to be omniscient.
Delegation needs boundaries
Specialist workers are useful when the task is bounded: review this repo, inspect this site, produce a patch, write a scorecard. They become risky when they hold the source of truth or decide what ships without a review gate.
What makes it operational
- Provider lanes are explicit.
- Repo paths and deployment habits are remembered.
- Verification happens with tools, not vibes.
- Background jobs report only what matters.
- Risky actions require a clear boundary or approval.
The point
A good AI stack should reduce repeated steering. It should not require the operator to become a babysitter for increasingly elaborate demos.
The control plane is where that discipline lives.