Build to compound
Keep people in charge of the agents you already run
You have several agents working. We give you the console, the approvals and the ledger, so somebody can answer what they did and who allowed it.
One agent is a feature. Six is an operation.
The questions change at about the third agent, and nothing built for the first one answers them.
Nobody knows what is running
Somebody asks what the agents are doing now, and the answer is assembled from three log searches and an impression.
Approval happened somewhere
An action went out. Which person allowed it, on what basis, and whether they saw what you are looking at now, is not recorded in one place.
The only lever is off
Something is going wrong and the choice is between watching it and stopping the service, because nothing between those was built.
Agent frameworks ship a runtime and assume somebody else builds the console, so the operational half arrives when somebody is already worried.
The shift
Three ways to supervise agents
Two of them stop working at about the third one.
Read the output
Fine for one agent
A dashboard of runs
Visible, not actionable
A mission console
Watched, steered, recorded
What we put in front of them
The operator surface, and the authority model underneath it.
Missions
Work framed as something with a goal and a boundary, rather than a stream of individual runs nobody can group.
Approval tickets
An action needing a person becomes a ticket with the agent's context attached, so the decision is made on what it saw.
A runs ledger
Every run, what it was given, what it did, who allowed it. One queryable record rather than four log searches.
Levers short of off
Pause a mission, redirect one agent, revoke a grant. The space between doing nothing and stopping everything.
Delegated authority
Who may approve what, scoped, so oversight is not one person becoming the bottleneck for every agent in the company.
Policy the agent cannot argue with
The gates are evaluated by the platform rather than requested in a prompt, which is the difference between a rule and a suggestion.
1 ledger
For every run, approval and intervention
Built when our own count passed the threshold
Tower exists because reading output stopped being supervision once we had several agents working at once. It is free and open source and it runs in your environment.
Clients we have done this for
Agentic GTM · Our own product
Canvass
Canvass brings account research and campaign review into a Cloudflare application. The workflow keeps campaign preparation, approval and delivery states distinct.
AI systems · Our own product
Fabric
Fabric is a context store, memory and agent runtime built on Databricks, and it is the clearest evidence of what we build: across our clientele, work that took a team of ten now takes three.
FAQ
Agent oversight: common questions
Do our agents have to be built on your framework?
No. Tower supervises agents rather than replacing the runtime under them. The tighter integration is with Harness because we built both, and it is not a condition.
Is this a compliance product?
It is an operations product that happens to produce what compliance asks for. Building it the other way around gets you a report nobody uses and no levers when something goes wrong.
How much can it actually stop?
As much as you give it authority over. The point of the delegation model is that the answer is a decision you make rather than one the tool makes for you.
If your problem is shaped differently
Build an agentic software factory
Launch a product in a regulated vertical
Take the AI demo to production
Move off the warehouse without a Friday night
Make the estate answer for itself
Modernise the estate that outgrew itself
Leave your team able to do it without us
See what production is doing, and be allowed to act