In-house tools · 2026
A control dashboard built by agents in one night
I asked an agent for a dashboard showing the rack’s status and usage per person. By the next morning it was up and running; then we polished it with three agents.
- for the first version
- 1 night
- metric resolution
- 5 min
- monitored nodes
- 3
- agents on the project
- 3
The brief
I wanted to see at a glance what each machine was doing and who was consuming each model. I asked Hermes’s web360 profile for it before going to bed.
What came out
- A collector that, every five minutes, reads the status of the two Sparks and the GPU server, plus usage from the gateway database.
- An API and a React web app with live status, a per-person ranking and usage by model, from the last few hours to the whole year.
- All in Docker, behind authentication.
What happened next
On day one, tokens were being double-counted and the ranking mixed up aliases belonging to the same person. OpenCode, with the coding model on the RTX 3090, wrote the patches, and Cursor checked the numbers against the gateway logs until they matched.
What I learned
An agent can build a useful tool in one night, but the numbers always get checked against the original source. The dashboard is now how I know, every day, that the lab is working.