I build environments where people and AI agents work on the same records under the same rules. These are three small demos of how I approach that.
Each one runs in your browser on made-up data. Nothing here comes from a real company's systems. Use the buttons at the top to jump to a demo.
1 · Bounded agents
A mock of the agent system I run at work
The graph the agent runs on
Each step is decided ahead of time to be run by a model, by exact logic, or by a person. Loops are allowed, and every loop has a limit.
The case
What is real here
At Print Reach I launched and still own an internal AI agent system that runs across four products and handles about 5,000 conversations a month. The ideas on this card are the ones it is built on: a graph engine with loops bounds each step, the context each step gets is chosen rather than left to the model, a person confirms the decisions that matter, and every action and its result stays on the case.
This page shows none of the real system. The graph, the requests and the replies are invented.
2 · One rule set
Schematically · open source, in beta
Ask for a connection
The same request goes to all three surfaces. Change it and ask again, or click two boxes in the diagram.
The rule, written once
The UI, the HTTP API and the MCP server all call this one function. None of them has its own copy of the rule.
In Schematically a diagram is a model, not a drawing: boxes, boundaries, ports and wires with rules about what may connect. People edit it in the browser and agents edit it over HTTP or MCP, and every surface enforces the same rules, so a person and an agent get the same answer. This card is a small reconstruction of that idea, not the editor itself.
In the editor person
Over HTTP program
—
Through MCP agent
—
3 · Grants and receipts
Soveraeign and my workstation · experimental
Define an agent's job
A person writes a contract: what to do, which files it may touch, how the result is checked, and the most the agent may claim.
Grants
An action is allowed only when a live grant covers it. Being able to do something is not permission to do it.
Task T-14
Receipts
Every attempt leaves one, allowed or refused. Each receipt includes the hash of the one before it, so history cannot be quietly rewritten.
Soveraeign is a local-first system where people and AI act on the same records, permissions and history. Authority comes from explicit grants, and every action leaves a receipt. I run my own work through a workstation built on these ideas, where agents take written contracts like the one above. It is experimental, and this card is a small reconstruction of it.