Addressable, persistent, bound to one asset
A general assistant is asked a question and forgets it. An asset agent is a standing entity in the graph: CH-02 has an agent, and that agent carries the chiller's design basis, its commissioning data, every intervention since, and the exceptions still open against it.
This matters because engineering context is cumulative. The reason a bearing failed in 2024 constrains what a sensible recommendation looks like in 2026 — and a stateless model has no way to know that.
Four capabilities, invoked on its own asset
The agent is the addressable entity; these are the functions it runs. Each operates over the graph rather than raw documents, and each cites the nodes it used.
Knowledge
Answers engineering questions about this asset with citations to the exact document revision and clause.
Failure mode
Ranks candidate failure modes against observed symptoms and this asset's own FMEA — never a generic library.
Intervention
Proposes inspection and maintenance timing, every option passed through PhysicsNET first.
Compliance
Checks that decisions and their evidence satisfy the standards and internal procedures in force for that asset class.
The limits are architectural, not configured
An agent that can be granted authority by changing a setting has authority. These constraints are properties of the system rather than options in it, so there is no permission an administrator can grant that lets an agent act on plant.
Blocked proposals are recorded too. A pattern of physics rejections is itself an engineering finding — usually about the asset, sometimes about the model.
Context, and what it is allowed to conclude from it
The last column is the discipline: what the agent read is fact, what the agent concluded is not. Both travel with every proposal.
Accountability does not move to the agent
Every asset has a responsible engineer, and every proposal is addressed to that person. Acceptance is a signed act with a timestamp; rejection is recorded with a reason and is fed back to the agent as context — never as training data for anyone else's model.
Addressed, not broadcast
A proposal goes to the engineer who holds the asset, not into a shared queue where ownership dissolves.
Rejection is information
The reason a competent engineer said no is the most valuable context the agent will ever receive, and it stays in the tenant.
The record survives the people
Who decided what, on what evidence, is retained for the asset's life — which is longer than most engineers' tenure on it.
Layer 06 of seven — above the gate, below the engineer
Asset agents sit above PhysicsNET and ENGIN Solver and below the engineering decision record. They are the layer that composes; they are not the layer that concludes.
Point an agent at one asset you already argue about.
Design-partner evaluations start with a single asset that has a contested history — the case where accumulated context either earns its keep or does not.