Your best operator, on every shift.

Your best operator knows that this AHU hunts when it is humid, and that the chiller that tripped last March tripped for a reason nobody wrote down. That knowledge currently leaves the building at 16:00, and permanently when they retire.

What an agent reads

The namespace and the record, and nothing else. An agent does not get a private data source or a separate view of the plant — it reads the same model your operators read, which is why its answer can be checked against the same evidence.

It explains, rather than alarming again

An alarm queue tells you a point went out of range. An agent tells you which piece of equipment is behaving wrongly, what pattern it matches, what changed shortly before it started, and what was done the last three times it happened. The alarm is an input, not the output.

The team cannot simply be bigger

Coverage is not a staffing decision anyone wins right now. In the US, the Bureau of Labor Statistics projects about 72,700 electrician openings a year, on average, over the decade, against 821,000 people employed in the trade in 2025 — and puts the apprenticeship that fills one of those openings at four or five years, with about 2,000 hours of paid on-the-job training in each year of it. The IBEW programme sits at the top of that range: five years and 10,000 hours. NECA, the contractors’ association, puts the average journeyman’s age at about 53, with close to 70% of supervisor roles held by Baby Boomers.

Those are US figures, and the pipeline arithmetic behind them is not: a trade that takes half a decade to qualify for cannot be scaled up inside a budget year anywhere. Buildforce’s assessment is that “no hiring strategy, no recruitment effort, and no wage increase can meaningfully expand the supply of journeymen in the near term.”

So the useful question is how much ground one qualified person covers: how many sites from the same desk, and which of the nights, weekends and holiday weeks nobody can staff. That is what an agent reading the namespace and the record changes. Not a smaller team — the same team, with the answer already assembled when they pick the job up.

Reach is granted, bounded and recorded

An agent may read what you let it read and write what you let it write, against the specific equipment you named. Anything outside that grant is not something it declines to do — it is something it has no path to. Every action it takes lands on the record beside the actions people took.

Specifics

Reads
The asset namespace and the retained record, scoped to the tenant
Writes
Only the equipment and actions named in its grant
Admission
Declared, tested and admitted against its named target before install

Common questions

What stops an agent acting outside what we granted it?

The grant is the path. An agent reaches only what its admitted manifest names, and a candidate that asks for more does not receive an edge directive at all. That admission process is the Oddy test, and its verdicts are retained so you can answer what ran and why.

Can an agent override a safety function?

No. Indagate is not a safety-rated controller. Safety relays, interlocks, E-stops and machine guarding are authoritative and sit outside anything an agent can reach.

What happens to the knowledge if we leave?

It exports with the rest of your record, on a documented schema. The knowledge compounds at your site because that is where it is useful, not because it is being held.

Bring your point list.

The first conversation is a survey: what you run today, and what it would look like modelled.