Shtab: a control desk for a team of AI agents

Task

The client wanted to bring AI into their daily work, and soon it was no longer a single model working for them but a team of agents. Terminals do not cope with that: you cannot see who is doing what, what is waiting for a decision, how much of the limit is used and where a risky step is. They needed one control desk: a window on a Mac that shows and runs all the agents’ work across every project, where risky actions (money, releases, secrets, messages to clients) do not go through without their “yes”.

What the client got

A macOS app with a fast core and a window that reads the agents’ logs and controls them.

• The World. Projects are shown as a city: each project is a tower, and agents sit on its floors. You enter an office with a click or by zooming in, with trackpad gestures like in Maps.
• Project director. One button, and the project gets a director: it reads the project’s rules and history by itself, assembles a team of agents with the right models for the task, keeps the roster and the journal.
• Project room. A chat with the director, channels and a decision feed in the familiar style of a messenger.
• Autopilot from a brief. Text, PDF, DOCX or a client conversation turns into criteria and a plan. The director works in rounds; Shtab wakes it, moves it to a fresh session when memory is full, and waits for the limit to reset. It stops only at gates: the plan, money, a new library, a release, a message to the client.
• Limits and a second account. A live limit counter per account, a forecast to the end of the window, switching tasks to a spare account, a Spending screen: who spends how much.
• Critics. An own critic on every merge, external reviews by other model families within a run budget, all in a sandbox with a temporary home.
• Sales. A lead funnel and first messages that only a person sends, by pressing a button (see the Leads project).
• Security. A threat model; a guard on agent steps that lets through only known commands and hides secrets from agents; a hash-chained journal in which tampering with a record shows; secrets only in the Keychain; a secret scanner before every commit.
• Assistant server. A separate product: the assistant reads client conversations, prepares drafts, proposes meetings for confirmation, sends reminders and replies from notes; a bridge for requests from the website; a conversation journal with a hash chain.

Result

A working control desk is built: the World and offices, a director and a team, decisions, sales, spending, the project room and autopilot from a brief.

Next stage: a Friends section for the team (being checked)

Now the client wants the whole team to use Shtab. For that there is a Friends section: a colleague is invited with an invitation or a code, people are gathered into groups, and everyone decides for themselves who sees what. You can open a colleague’s World, see what their agents are working on and message them right away. Agents’ tasks and reasoning are visible only if the colleague has opened the “Detailed” level; at the “View” level only the overall picture is visible. The section is already built and is being checked now, but it is not released in the app yet. The shots in the gallery show the real screen of the section, and the data on it is made up.