Solar and green-power agent
In solar and green-power scenes the model focuses on multi-scenario reading and report structure; the harness locks irradiance, output, prices and carbon methodology inside allowed bounds, pulls forecasts and financial parameters only through allow-listed tools, and emits versioned, evidenced drafts — so a verbal IRR cannot go unchecked.
Under irradiance, offtake constraints and grid-connection rules, fuse output forecasts with power-price and carbon-price signals to produce multi-scenario IRR and abatement per kWh — materials for plant expansion and green-power contract talks.
Combine satellite remote sensing, shading and available-site data to estimate installable capacity and annual generation bands, and flag the marginal effect on park carbon intensity and green-power share — so electricity is never calculated without carbon.
Operating loop
From connection alignment to review versions, every step can be checked, trailed and approved.
Connect and align
Pull array and grid-connection metering and status from the time-series store and asset ledger; align to a shared clock and missing-data rules; block later reasoning and prompt for backfill if the quality gate fails.
Inject goals and constraints
Write the current decision goal (expansion, recontract or offtake lift) and hard constraints: interconnection capacity, offtake cap, contract form. The harness structures these as model context so conversation cannot quietly change methodology.
Controlled tool calls
Only pre-registered forecast, price-curve and simple financial trial tools may be called. Each call checks parameter range and tenant permission, rejects overreach or unknown tool names, and logs request and response as a step.
Draft generation
The model organises conclusions across scenarios: generation bands, bill and Scope 2 sensitivity copy. Output must be structured fragments (paragraphs + table placeholders) for later template render and diff.
Guardrail review
Rule-check the conclusion: factor version vs organisation library, locked price path cited, unauthorised commitment language. Failures roll back for rewrite or are marked for human edit.
Human approval gate
Outputs that become external materials or cross investment thresholds enter an approval queue. Approvers can inspect the tool-call chain and data slice; only then may they export or enter collaboration.
Review and version
The final draft carries data window, factor and model version. Daily/monthly look-back of “what we relied on then” supports internal control and later strategy comparison.
Typical tool capabilities
Callable inside the allow-list
Guardrail points
Bounds the model cannot override
- Interconnection and offtake constraints cannot be overridden by the model
- Emission factors and organisation boundaries are read-only organisation-library versions
- Tool allow-list and per-call cost/count caps
- External documents forbid unauthorised commitments and absolute-return claims
- Tenant- and station-level data isolation
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.