Human oversight console
Keep people in charge of the agents you already run
Mission control for AI agent squads. Where people watch, steer, and approve the work.
- Role in the family
- Human oversight console
- Databricks surfaces
- GenieUnity AI GatewayDatabricks AppsLakebaseTemporal
- Status
- Phase one is live: a real squad runs end to end. The durable path is certified against Temporal Cloud: a mission parked at approval, survived a full application restart, and completed from its original workflow history.
- Documentation
- tower.fabric.pro
The problem
Once several agents are working, the question stops being 'can it run' and becomes 'what is it doing, who approved that, and how do I stop it'. Most agent stacks have no answer because they were never designed for a human to supervise.
How it works
A mission board, one-tap approval tickets, a runs ledger, and shared documents, with the governance pipeline under every mutation. One request plans through a supervisor agent and dispatches to the squad. Anything that would leave the workspace parks at a ticket. The supervisor never approves its own work; a person does.
What it does
One ask runs the squad
A request is planned by a supervisor and routed per task to the right specialist.
Humans hold the gate
Anything leaving the workspace stops at a ticket a person answers. The supervisor cannot approve itself.
Genie-backed analysis
Data questions route to a Genie-backed analyst; the answer and its SQL are the deliverable.
Survives restarts
Approval-parked missions resume from workflow history after a full stop and start.
Your people keep authority over what the agents do. A mission board, approval tickets, and a runs ledger give operators one place to watch, steer, and stop the work. It is the control surface most teams find they need only after something has gone wrong.
Several agents are working. Nobody is watching.
One agent is a feature you can supervise by reading its output. Six agents is an operation, and the questions change.
What is running right now
Somebody asks what the agents are doing and the answer is assembled from three log searches and a guess.
Who approved that
An action went out. Which person allowed it, on what basis, and whether they saw the same thing you are looking at now, is not recorded anywhere one place.
There is no way to intervene
Something is going wrong and the only lever is stopping the whole service, because nothing smaller was built.
Agents get built with a runtime and no console, so the operational half arrives only when somebody is already worried, and by then it is being built under pressure.
The shift
Three ways to supervise agents
Two of them stop working at about the third agent.
Read the logs
Fine for one agent
A dashboard of runs
Visible, and not actionable
A mission console
Watched, steered, and recorded
What gets installed
The operator surface that agent frameworks assume somebody else will build.
Missions
Work framed as something with a goal and a boundary rather than as a stream of individual runs nobody can group.
Approval tickets
An action needing a person becomes a ticket with the context attached, so the decision is made on what the agent saw rather than on a summary.
A runs ledger
Every run, what it was given, what it did and who allowed it, in one queryable record.
Squad oversight
Several agents working together, visible as a group, with the dependencies between them legible.
Intervention short of stopping
Pause a mission, redirect one agent, revoke a grant. Levers between doing nothing and turning it off.
Delegated authority
Who may approve what, scoped, so oversight is not one person becoming a bottleneck for every agent in the company.
1 ledger
For every run, every approval, every intervention
The console we needed once we had several
Tower was built when our own agent count passed the point where reading output was supervision. It is free and open source under Apache-2.0 and it runs in your environment.
FAQ
Fabric Tower, answered
What does Tower actually give an operator?
One place to watch, steer and stop the work of an agent squad: a mission board, one-tap approval tickets, a runs ledger and shared documents, with the governance pipeline under every change.
Can an agent approve its own work?
No, and that is enforced structurally. Anything that would leave the workspace parks at a ticket that a person answers. The supervisor plans and routes; it never approves.
What happens if the system restarts mid-mission?
The mission resumes. The durable path is certified against Temporal Cloud: a mission parked at approval survived a full application stop and start, then completed from its original workflow history.