promptishInside Promptish
How we run our own operation on AI agents
An internal case study from Promptish. This describes our own setup, not a client engagement. Facts reflect our operating records as of November 1, 2026.
A shared record for the work
Claude leads the work. Annie and Kobe are two Codex seats that take on assigned tasks. They coordinate through a shared, Git-backed operations repository containing inventory, decisions, plans, and runbooks.
That gives the agents a common place to look up what exists, what was decided, and what still needs doing. The rule is simple: record decisions, and treat untested work as untrusted.
Inventory instead of guesswork
Automated inventory covers three servers, including one supporting a 24/7 automated workload. It records hardware, services, timers, and failures. The inventory schedule is configured for every six hours.
Those records show problems as well as running services. A service appearing in an inventory is not proof that the whole workflow works correctly.
Local AI for background work
We are setting up a self-hosted local language model on a CPU server for background tasks such as inquiry triage. The model service is configured for local access with limits on CPU and memory use.
The decision log records model downloads in progress and benchmarking still pending. We do not yet claim measured speed, quality, or cost savings from this setup.
A queue with approval rules
The dispatch queue records proposed, reported, applied, refused, and escalated work. Our operating rules require time and spending caps for tasks, and operator sign-off for money movements, contracts, new accounts, and public commitments.
Review gates and automatic spending controls are part of the dispatch design. The inventory still lists several of those safeguards as planned, so we do not claim that every gate or cap is automatically enforced end to end.
What this shows
We use AI agents for real internal work, with shared records and human responsibility. We are still testing the system and closing gaps. This case study makes no claim about client outcomes or measured time savings.
The same approach shapes our agency offer: choose a specific task, agree what a useful result looks like, and keep approval with the people responsible for the work.