UMAP · Universal Multi-Agent Processing
Whether an agent tracks attrition, detects fraud, routes baggage, predicts readmission or scores supplier risk, it is structurally the same thing. It has the same layers, the same kinds of component, and the same shape. Every agent. Every domain.
If that is true, then building a new AI system stops being a construction project and becomes a configuration exercise. Everything else in the framework follows from that one idea.
Every agent carries the same six. What changes between domains is the value in each, never the structure.
What the agent is, who owns it, and what it is allowed to decide on its own.
What wakes the agent up, under what conditions, and how often it is permitted to fire.
How it reasons, how sure it has to be before acting, and what it remembers.
What it does with its conclusion, where a human stays in the loop, and what happens when it cannot resolve.
How human decisions flow back in, and how the agent improves because of them.
How the agent is seen, measured and governed once it is running.
No component state is ever overwritten. Each change writes a new dated record, so the full state of any agent at any past moment can be reconstructed.
What was this agent doing on the day in question? Answerable from the record, not from memory.
When did a threshold change, and why? Every change is a dated entry with an owner.
What state was the agent in when the error hit? Reconstructable exactly.
Restore any agent to any earlier state without a redeployment.
It is a structural property of the model. It costs nothing to design in, and everything to design out — which is why most agentic estates cannot answer a regulator's question about last quarter.
Five consequences, each of which changes what a build costs.
Because the object is already defined, building in a new domain means filling in what is domain-specific and inheriting everything else.
Algorithms live in a shared library and are referenced, not copied. There is no per-client model to maintain.
Roles map to pre-built personas, so a new user's agents, views and thresholds are ready on first login.
Agents read from a prepared feature layer rather than from source systems, so the plumbing can change without the agents changing.
Not because corners are cut — because the framework removes every decision that does not need to be made a second time.
UMAP is not a product you buy. It is the reason the products behave the way they do.
Every packaged product is an instance of the same object model, which is why the second client in any industry starts from the first client's engine. See the catalogues.
The build layer stands agents up against this model rather than against a bespoke design each time. See BackX.
Supervision is possible because every agent exposes the same observability surface. See TechnologyX.
The 802-block checklist is what a system needs around this model to run in production. See the checklist.
The component specification, the addressing and versioning scheme, the instantiation method and the library reference model are the subject of a pending patent application and are not described on this page. What is above is the shape of the idea. The rest is covered under NDA in a technical session.
A ninety-minute session produces the agent map, the data products behind it, the persona mockups and the ROI framework — for your process, in your language.
Book a ninety-minute session