Microgrid and park agent
Microgrids and parks emphasise real time and safety. The harness confines charge/discharge, demand and backup strategy to physical and contract-allowed bands; the model narrates trade-offs on forecast curves and ranks options. Real dispatch requires strategy simulation and signed strategy packs.
In source-grid-load-storage scenes, schedule charge/discharge and demand management on 15-minute-to-day rolling forecasts, balancing backup and peak–valley spreads; map dispatch results to carbon and bill dual metrics for operations reviews.
Under load growth, distributed capacity and quota constraints, rank storage/PV expansion and give a rough investment band, explaining the path to purchased-power cost, green-power offtake rate and Scope 2 emissions.
Operating loop
From connection alignment to review versions, every step can be checked, trailed and approved.
Situational alignment
Aggregate real-time and near-real-time source-grid-load-storage measurements; align clocks and topology; tag planned outages, spot swings and other external events into loop context.
Rolling-forecast pull
Pull load, output and price forecasts at 15-minute-to-day scale. The harness checks forecast-service version and validity window; stale forecasts trigger recompute or degrade to a conservative strategy.
Optimise and simulate
Inside guardrails, call dispatch-simulation tools to generate charge/discharge and demand options; compare bills, backup coverage and Scope 2 emission proxies into a comparable indicator table.
Strategy draft and explanation
The model translates the preferred option into operating notes and meeting language, and lists bound assumptions (spreads, load error bands). Unverified device command strings are forbidden.
Strategy-pack checks
Hard-check setpoints, ramp rates and SOC bounds against interconnection agreements and demand contracts clause by clause; conflicts auto-reject or require human confirmation.
Dispatch and trip
Approved packs go out on a controlled channel or become work orders. On runtime faults (comms loss, limit breach) the harness runs a preset degrade or stop template and records a reason code.
Operations review
Archive actual vs plan deviation; produce a daily “bill–carbon” dual-metric review summary as prior weights for the next loop.
Typical tool capabilities
Callable inside the allow-list
Guardrail points
Bounds the model cannot override
- SOC, power and ramp hardware bounds
- Point-of-connection power and power-quality constraints
- Demand-contract and spot-exposure caps
- No unauthorised direct control commands
- Key actions require dual approval or role permission
Other domain agents
One harness control plane, scene-specific guardrails
Put the agent inside a governable loop
A demo covers tool bounds, approval gates and how to assemble it privately.