CHAPTER 03 · BUILD
Agents Shipped Inside the Existing Stack
The first system, and the one right behind it, shipped inside the control room's existing screens — no new SCADA integration, no rip-and-replace. Each one is real and running.
THE AGENTS · 3 FOR THIS WORKFLOW
This control room needed two. The count isn't a package — it's however many distinct jobs the work splits into. Each agent is a separate identity with its own scoped tools and permissions.
Grid Anomaly Monitoring & Explanation
Generation-Asset Risk Monitoring
Dispatch Prioritization
IT DOESN'T STOP AT A REPORT
The Agent Takes the Action
An explanation is one output. The agents also take the next action — but on a live grid, a human confirms anything that touches the network, every time.
Explain the anomaly
Writes the plain-language cause next to the alarm, cited to the telemetry — not just a red light.
Draft the corrective action
Prepares the switching or dispatch step for the operator to confirm — nothing touches the grid without a human.
Notify the right desk
Routes the event to the correct control-room role the moment it's classified.
Log the event and its cause
Writes the event, its cause, and the action taken into the record automatically.
PRODUCTION TRACE
Grid Anomaly Monitoring & Explanation
Every step below runs against the plant's real data — the exact sequence, in order.
7
Anomalies explained
6
Confirmations accepted
1
Escalations sent up
2
Shift-handoff notes carried forward
2 items need your judgment
Ingest
AutomatedLive telemetry pulled from 7 balancing-authority feeds every 5 minutes via the real-time EIA API — no new sensors, no new data entry.
Detect
AutomatedA rolling statistical model flags deviations against each authority's own recent baseline — not a fixed threshold that's wrong half the time.
Explain
AutomatedEvery flagged anomaly gets a plain-language explanation — what changed, which authority, how it compares to the last 30 days — not just a red light.
Operator review
Reviewer inputEach explanation is one-click confirm or escalate. Escalations become training signal for tomorrow's detection.
Shift handoff diff
AutomatedThe next shift opens with a diff against today's confirmed anomalies — did it recur, what changed.
REAL OUTPUT · GENERATION-ASSET RISK MONITORING
One Example, Fully Processed
When a job calls for a written artifact, this is what the system produces — everything it finds, structured for the action it triggers next.
Unit 4 — Combined-Cycle Generation
NRC/EIA-style real-time monitoring · Evaluated every 5 minutes · Live data, local model
Generation-Asset Signals Evaluated
| Signal | Reading | Baseline | Unit | Confidence |
|---|---|---|---|---|
| Output deviation | -6.2% | vs. 30-day avg | Unit 4 | 94% |
| Heat-rate drift | +3.1% | vs. commissioning spec | Unit 4 | 88% |
| Vibration signature | Nominal | vs. baseline | Unit 2 | 97% |
| Emissions reading | Nominal | vs. permit limit | Unit 4 | 99% |
| Startup-cycle count | +2 vs. plan | vs. maintenance interval | Unit 3 | 71% |
Fleet Summary — for the Morning Brief
Unit 4's output deviation and heat-rate drift are the two signals worth a look today — both consistent with a compressor-fouling pattern the fleet has seen before, not an immediate reliability risk. Units 2 and 4's other signals are nominal. Unit 3's startup-cycle count is running ahead of plan, worth a maintenance-interval check.
Watch List — Engineering Briefing
WAVE 2 — IN BUILD
Dispatch Prioritization — What's Next
When a regional demand spike hits during a generation-unit derate, dispatch has to decide which units to ramp and which loads to prioritize — today that call is made from memory and a phone tree, not a documented tradeoff.
The agent proposes a dispatch order; a constraint solver checks the math against real capacity and ramp-rate limits. Nothing ships on an LLM's arithmetic — the solver's answer is the answer, the agent just explains it.
Worked example the solver evaluates today in testing
A 4% regional demand spike hits while Unit 4 is derated 15% for the heat-rate issue above. Four dispatch options scored by projected reserve margin:
WHAT CHANGED
Time returned
~85% of manual triage time
Operators got the time back for judgment calls, not for piecing together what an alarm meant.
Operating complexity
A ten-minute scramble across three systems became a one-click-confirm explanation
Where this stands
First system live in 21 days
UP NEXT · TEAM VIEW