The firm had agents spread across dozens of apps, each built separately. When one went quiet, nobody knew. When a prompt needed changing, someone redeployed an app. Nobody could say what an agent cost per day.
Four people needed four different things: operators wanted outages and runaway spend, prompt engineers wanted to edit without a deploy, finance wanted real token cost, and developers wanted a fast way to wire a new app in.
What an operator sees, in the order they'd use it.
Registered apps push heartbeats and run reports. Each agent's status becomes running, idle, paused, stopped or error, and a rolling health score recalculates on every run. Agents can have parents, so hierarchies show up as hierarchies.
Pause, resume, stop, restart, swap the model or push a new system prompt. Each command is an HMAC-signed webhook to the live app, retried with backoff and tracked from pending to acknowledged.
Every prompt keeps its full history with one-click rollback. Feedback folds straight in with tags, and an AI reviewer scores answers from 0 to 100 with a written reason.
Its 8 internal AI features run as real tracked agents, from the integration analyzer to the error solver, so platform spend is measured with the same meters as everything else.
Every heartbeat, report and command lands in one stream. Fallbacks get an amber badge, so an operator is never told the primary model answered when it didn't.
The feed on the right is a live sketch. It is generated in your browser, not pulled from real data.
One Node process serves the JSON API and the compiled React app, split by module rather than by process. No broker, no worker fleet. A tool that watches moving parts should have as few of its own as possible.
Ask a free, deterministic public price API first. Only if that fails do two AI searchers research in parallel. If they disagree, each sees the other's answer and re-checks; if they still disagree, a third model judges. Everything goes to an audit log, with per-model cooldowns and a global concurrency cap.
Because a stopped agent has to stay stopped, even if its app keeps reporting in. Token counts are upserted atomically too, so two concurrent heartbeats can't lose a count.
A production firewall read legitimate prompt text as an attack and blocked saves. Base64-encoding the prompt and feedback write paths got past it without weakening the firewall.
Need something like this? I build these end to end, as a full-time engineer, on contract, or with SKAI Group.
Book a 30-min callRunning agents nobody can see? This is the problem I like most.