The method
One framework. Five steps. Every step leaves something you keep.
Agents only work when the ground under them is solid. The Loom is how we make it solid — one governed path from “is this environment ready” to “this agent runs, and we can prove it.”
Score it
Before any agent is built, we score the environment it will act in. Where do the same facts live in three places? Which data is stale? Who owns which rule?
You keep: a scored map and a fix-first listSettle the truth
One system is named the owner of each fact — worker, time, credential, pay — and we write down how the systems stay aligned. Disagreements between systems get settled once, on paper, instead of every time an agent asks.
You keep: a source-of-truth hierarchy and alignment contractsWeave one picture
The systems are joined into a single, current picture an agent can read in one call. Every fact in it is checked at the moment of decision — duplicates collapsed, conflicts resolved by the ownership rules, each fact scored for confidence and traced to its source.
You keep: one validated model your agents read in a single callEarn autonomy
The agent starts by answering questions. Then it acts on command. Then — action by action, only as the checks pass — it owns an outcome. Autonomy is earned. Consequential actions keep a human confirmation where policy requires it.
You keep: a tiered agent design with explicit graduation criteriaProve it, daily
Every decision a live agent makes is scored: was it right, was it grounded, was it within policy, is the outcome holding. When anything drifts, the finding goes back to Step 1. The loop closes, and the next agent starts from a stronger foundation than the last.
You keep: a live trust scorecard and an audit trail a regulator can followOne scenario, end to end
A rule that touches three systems and lives in none
A hospital’s rule: a nurse charging time to a sedation case must hold a current sedation credential. The rule depends on the worker system, the time system, and the learning system — and lives in none of them.
Step 1 finds the credential kept in three places, one of them eighteen hours stale. Step 2 names one owner. Step 3 turns the rule into one picture: this nurse, this credential, this shift — current, checked, one read. Step 4 walks the agent from answering “is she credentialed?” to owning the outcome: every sedation case for the next two weeks staffed by a credentialed nurse, gaps fixed before the case.
Step 5 measures one number — cases that reach the floor without a valid credential — and keeps it at zero.
An illustrative scenario — how the method would run in a multi-system workforce estate.
The loop compounds
The same five steps run on the next use case, and the foundation built for one agent makes the next one faster. That compounding is the asset.
REIMAGINEBUILDOPERATE