← All case studies

Case study 03 · Operator dashboards

AI agents are only useful if a person can see and steer them.

We built two control rooms for our own AI workforce. The first taught us what a dashboard gets wrong when it trusts its inputs. The second fixed it: every screen state comes from a recorded event, and nothing risky happens without a human saying yes.

13agent roles tracked by the controller
3parties in one room: you, Claude, Codex
30.4sfull live acceptance run, restart included
0risky actions without human approval
The problem

Agents that work in the dark are a liability

Once AI agents start doing real work, three questions matter more than any model benchmark: what are they doing right now, what did it cost, and who approved it? Most teams find out after something has already gone wrong.

System 1 · Factory Controller

One screen for apps, builds, agents and spend

The first control room ran the App Factory: every app and its status, one-click Android and iOS builds, a full build history, each agent run with its cost, and App Store submissions and rejections. Everything persisted in a local database, so nothing was lost between sessions.

Interface recreated for this page with demo values. No client data.

What it taught usThe controller faithfully displayed whatever its agents reported, including results that were confident and generic rather than checked. A dashboard is only as honest as the events feeding it. That became the founding rule of the second system.
System 2 · Command Deck

A control room where people and AI agents work together, on the record

Command Deck is one room for three parties. The human is the Director. Claude answers first as Strategist; Codex then sees the question plus Claude's unedited answer and responds as Builder and Verifier. You can address both, or one with @claude or @codex, and a structured debate is capped at two rounds.

Interface recreated for this page. Conversation is illustrative.

  • Every screen state comes from a recorded event. No fake typing, no invented "working" animation. If the log doesn't show it, the screen doesn't claim it.
  • Humans approve anything risky. External writes, money, credentials, destructive actions, production changes and public releases always wait for the Director.
  • Safe by default. New rooms start in Observe mode; Build mode is switched on deliberately. It runs only on the local machine, and secrets are never stored.
  • Scoped memory. Agents read project context through a read-only gateway that can only see the current room, with sensitive text redacted.
Prove

Accepted on real agents, not demos

On 30 August 2026 the live acceptance suite ran real Claude and Codex replies through the full cycle: a message to both agents, an exact Claude-to-Codex handoff, a server restart and a resumed session. It passed in 30.42 seconds, with no writes to protected files and redaction intact.

Keeping it realThe animated office view, Build-mode approval exports and Mac packaging are still in progress. What's shown above is the working model and a recreation of the interface, not a finished product for sale.
What it means for you

If AI touches your operations, you get a control room

  • A live view of what every automation and agent is doing, and what it costs.
  • Approval gates wherever money, customers or production systems are involved.
  • An audit trail you can hand to a client, an auditor or your board.
Automation without a control room is just a faster way to make mistakes nobody sees.

Related reading: Why we start every engagement with a five-day sprint