Humans and agents · Part 6 of 6

Modeler Agent guide: prompts, typed actions and safe handoffs

How humans and AI agents should operate Modeler: reference the active scenario, make validated edits, run typed workflows, verify action chips and return reproducible evidence.

For: Operators, analysts, AI agents and tool integrators
Time: 45 minutes plus any requested solve
Updated: 20 Jul 2026

Typed workflows keep Agent actions connected to named inputs, scenarios and persisted results. · Hosted on Cloudinary.

Direct answer

The Modeler Agent is safest when a prompt names the target asset or selected component, the value and unit, the active scenario, the action, the comparison and the required truth label. Every mutation should appear as a typed action chip and be verified in the model afterward.

Solved result first

What the completed analysis must expose

Open the technical text / video script →

An AI agent should operate Modeler as a typed, testable toolchain. The acceptance benchmark is not eloquence: it is exact target resolution, validated units, explicit scenario scope, an inspectable action chip, read-back verification, immutable result IDs, matched comparisons and a challenge step that tries to falsify the conclusion. Agents should return compact evidence packets that another authorised agent or engineer can replay.

Agent execution contract—minimum evidence for a completed technical task

Gate	Required output	Pass benchmark	Reject when
Context	Project, model, scenario, selected asset	All four visible or explicitly null	Prompt relies on hidden or stale state
Mutation	Typed action with old/new value and unit	Exactly one intended target; base unchanged	Partial-name guess or free-form database write
Data	Dataset ID, normalized scope, unit, lineage, receipt	Replayable without a natural-language guess	‘Latest data’ with no window or geography
Solve	Formulation, horizon, stride, status, result ID	Matched settings for every compared run	Only a screenshot or completion toast
Analysis	Absolute values, deltas, mechanism and benchmark	At least one physical/economic cross-check	Headline delta with no denominator or cause
Handoff	Sources, assumptions, limitations, next test	Another authorised operator can reproduce it	Secrets, restricted raw rows or unsupported claims

Analytical method

01Make workspace state explicit, then write a prompt with target, value, unit, scenario, action, comparison and evidence request.
02Prefer typed first-party tools for reads, edits, Data Shop pulls and multi-step workflows; refuse ambiguous asset matches.
03After every mutation, read back the effective value in Tables or the scenario override surface before solving.
04Require result IDs, solver settings, modeled horizon, absolute values, deltas and asset-level mechanism in the answer.
05Challenge the result with a unit check, balance identity, benchmark and sensitivity; record what could falsify the conclusion.
06Export a minimal replay packet that excludes credentials and restricted data but preserves every decision-critical reference.

Who should interrogate what

Electrical engineer

Can the agent point from KPI to component?

Require the bus, line, generator, load or storage ID and the constraint or time step driving the result.

Environmental engineer

Can it distinguish total, average and marginal impact?

Ask for source window, temporal alignment and the dispatch mechanism behind carbon or curtailment changes.

Energy economist

Can it state the counterfactual and excluded cash flows?

Require matched cases, units, denominator, objective and a list of missing capex/commercial inputs.

Policy maker

Can the claim survive outside the prompt?

Ask for scope, jurisdiction, assumptions, distributional effect and which evidence would be required for compliance.

Energy trader

Can it separate model proxy from market product?

Require product, delivery window, price unit, source and whether the output is observed, derived, forecast or modeled.

Before you start

• An active project, model and scenario.
• A selected component when using phrases such as ‘this node’ or ‘the selected line’.
• Permission to make the requested model change or launch the requested study.

What you will have

You will be able to write unambiguous prompts, verify typed edits and workflows, export the transcript, and hand another human or agent a compact reproducibility packet.

Step-by-step

Complete the workflow

Set the workspace context first

The Agent receives the active model, scenario and selected component context; make that state visible before asking for an action.

Modeler workspace with active project, network, component tree, scenario and Agent panel visible together — Confirm project, model and scenario context before asking the Agent to read or change anything. · Captured in Modeler and hosted on Cloudinary.

Do this

1.Activate the intended project, model and scenario.
2.Select the exact bus, line, generator, load or storage unit when using a pronoun.
3.State whether the base may be changed; default to a clone for scenario work.

Expected evidence

✓ The prompt and the visible workspace point to the same target.

Trust check

The Agent refuses ambiguous partial names; that is a safety feature, not a failure to be bypassed.

Use the six-field action prompt

A strong prompt contains target, value, unit, scenario, action and verification request.

Modeler map with a bounded Agent request drafted beside the active India network and scenario — A strong prompt names the target, value, unit, comparison and evidence expected from the Agent. · Captured in Modeler and hosted on Cloudinary.

Do this

1.Name or select the component.
2.Give the numeric value and physical unit.
3.Name the scenario or request a clone.
4.Say whether to edit, solve, compare or explain.
5.Ask for the action summary and result IDs.
6.State the claim boundary.

Expected evidence

✓ A typed tool call with explicit arguments instead of a vague prose answer.

Copy into Modeler Agent

On scenario ‘BESS 30 MW · 4 h’, set the selected storage unit p_nom to 30 MW and max_hours to 4 h. Show the validated edit action, then read back the component name and saved values. Do not change the base scenario.

Trust check

Never ask an agent to ‘make the result better’; specify the permitted input change and the metrics to compare.

Start with a read-only evidence query

Before authorising mutation or compute, make the Agent summarize the active context and identify missing evidence.

Modeler workspace with GridAI open for a read-only prompt that requests active project, model, scenario, component counts, horizon and latest result without mutation — Require the Agent to restate the full active context and available evidence before authorising a model change or solve. · Captured in Modeler and hosted on Cloudinary.

Do this

1.Ask for active project, model, scenario, component counts, horizon and latest compatible result ID.
2.Ask which requested metrics are already persisted and which require a new solve.
3.Require exact component matches and units for any asset referenced by name.
4.Ask the Agent to list one ambiguity or missing input that blocks a stronger claim.

Expected evidence

✓ A read-only answer grounded in visible workspace state.
✓ No mutation, Data Shop ingest or solve launched during context discovery.

Copy into Modeler Agent

Read the active Modeler context without changing anything. Return project, model, scenario, component counts, modeled horizon, latest result ID, available analytics and the single most important missing input for a defensible comparison. If any asset name is ambiguous, stop and list the matches.

How to read the result

• A read-first pattern prevents the Agent from solving the wrong scenario or duplicating a persisted run.
• The strongest agent response distinguishes available evidence from work it proposes to perform next.

Trust check

Do not grant mutation authority to an agent that cannot correctly restate the active context.

Verify parameter edits through the validated override path

Generic component edits are validated against the shared attribute registry and stored as scenario overrides.

Modeler Tables view used to verify component parameters after an Agent-proposed model edit — Review the exact component, field, old value, new value and unit before accepting an edit. · Captured in Modeler and hosted on Cloudinary.

Do this

1.Review the action chip for component type, ID/name, field, old value and new value.
2.Open Tables and verify the effective scenario value.
3.If multiple names match, provide the full asset name or select it in the UI.

Expected evidence

✓ The edited value appears in the intended scenario and nowhere else.

Trust check

Do not use raw database mutation or an unvalidated free-form field when a typed attribute exists.

Use first-party workflows for multi-step studies

Typed workflows are safer than asking the Agent to improvise a long mutation-and-solve chain.

Modeler specialist-study dialog describing source-hashed inputs, immutable runs and matched solver settings — Prefer a typed workflow when the operation needs repeatable inputs, persisted outputs and comparison gates. · Captured in Modeler and hosted on Cloudinary.

Do this

1.Use run_bess_study for base solve → rank → clone → place → re-solve → compare.
2.Use the Dholera data-centre workflow for the named Gujarat load and optional BESS study.
3.Review all workflow inputs before accepting the run.

Expected evidence

✓ A persisted workflow result with named clones and paired result IDs.

Trust check

A workflow implements a specific contract. Do not describe it as SCED, AC OPF or investment optimization unless that contract says so.

Give agents Data Shop receipts, not vague dataset names

A reproducible data instruction includes dataset ID, state list, from/to, resolution, target and model ID.

Data Shop configuration opened from Modeler Agent context with active-model attachment controls — The Agent may guide a data pull, but the user still verifies lineage, scope, unit and destination. · Captured in Modeler and hosted on Cloudinary.

Do this

1.Copy the Data Shop ingest receipt into the task context.
2.Include lineage and unit beside the scope.
3.Require the agent to restate the normalized scope before using it.

Expected evidence

✓ A data reference that another tool-capable agent can replay without guessing.

Copy into Modeler Agent

Use dataset iex_price_band with states [gujarat], from 2025-04-01, to 2026-04-01, resolution state, target attach and the active model ID. Treat the materialized unit as a derived-B 0–1 index, never raw ₹/MWh.

Trust check

A friendly name such as ‘latest Gujarat prices’ is not a reproducible data contract.

Require evidence-complete answers

Every result answer should include source windows, assumptions, model scope, solver/horizon, result IDs, key deltas and limitations.

Modeler Run evidence workbench beside GridAI with the persisted run, provenance fields and a bounded verification prompt visible — Check the persisted run, source inputs, units and tolerances behind an Agent summary before repeating or sharing its numbers. · Captured in Modeler and hosted on Cloudinary.

Do this

1.Ask the Agent to group observed, derived, assumed and modeled items.
2.Require absolute values and deltas with units.
3.Ask what missing input is most likely to change the conclusion.

Expected evidence

✓ A compact decision answer whose claims can be traced to the workspace.

Trust check

If the answer omits a result ID, modeled horizon or truth label, treat it as an explanation—not a completed study handoff.

Verify the Agent answer against Analytics and Results

Treat the narrative as an index into persisted evidence; independently verify every decision-critical number and denominator.

Modeler Results workbench beside GridAI for checking the Agent answer against the immutable run, units, horizon, arithmetic and asset-level evidence — Verify every decision-critical Agent number against the product source of record and recompute at least one ratio independently. · Captured in Modeler and hosted on Cloudinary.

Do this

1.Open the cited result ID and compare absolute values, deltas, units, horizon and solver status with the Agent answer.
2.Recompute at least one ratio, such as gCO₂/kWh, percentage cost change or BESS round-trip efficiency.
3.Trace one KPI to an hourly dispatch, component table or affected corridor.
4.Check that no modeled price is relabeled as an IEX observation, tariff or forecast.

Expected evidence

✓ Every quoted number matches the persisted result or an explicitly shown calculation.
✓ At least one independent arithmetic and one physical-mechanism check pass.

Worked calculation

Verification ratio examples:
Δ% = (scenario − base) / base × 100
Carbon intensity = emissions tonnes / energy GWh = gCO₂/kWh
BESS RTE = ηcharge × ηdischarge

How to read the result

• Agents are good at synthesis but can still omit denominators, mix horizons or repeat stale results; the product surface remains the source of record.
• A correct number with the wrong unit or time horizon is still a failed verification.

Trust check

Do not forward an Agent result until the immutable run, arithmetic and physical mechanism agree.

Ask the Agent to falsify its own conclusion

A trustworthy agent should name the sensitivity, benchmark or missing evidence most likely to reverse the recommendation.

Modeler Analytics beside GridAI with a falsification prompt requesting the sensitivity most likely to reverse the current conclusion — Ask for a bounded low, base and high test, an explicit decision threshold and the smallest run that could falsify the recommendation. · Captured in Modeler and hosted on Cloudinary.

Do this

1.Ask for the strongest counterargument to the current interpretation.
2.Ask which single input has the highest decision sensitivity and what range is plausible.
3.Require one alternative mechanism for each important KPI delta.
4.Ask for a bounded next run that could falsify the conclusion, not a vague request for more analysis.

Expected evidence

✓ A low/base/high test with exact inputs and unchanged comparison controls.
✓ A clear failure threshold and evidence request.

Copy into Modeler Agent

Challenge the current conclusion. Identify the one plausible input or model limitation most likely to reverse it, define a low/base/high sensitivity with units, say which metrics should change and why, and propose the smallest new run that could falsify the recommendation. Keep the existing base unchanged.

Agent challenge benchmark

Check	Weak answer	Pass answer
Counterfactual	‘Results may vary’	Named input, range and expected mechanism
Sensitivity	Many simultaneous changes	One bounded change with matched controls
Threshold	No decision boundary	Metric value at which recommendation flips
Evidence	‘Get more data’	Named source/field and why it resolves uncertainty
Reproducibility	No run settings	Scenario, horizon, solver and outputs specified

Trust check

Confidence should increase because the conclusion survived a targeted challenge, not because the Agent used assertive language.

Export the transcript and minimal agent packet

A handoff packet should be small enough for an agent context window and complete enough for reproducibility.

Modeler result workspace with Download control and a structured Agent prompt for a reproducible review handoff — End with exported evidence, source provenance, assumptions, limitations and the next action for a reviewer. · Captured in Modeler and hosted on Cloudinary.

Do this

1.Download the chat transcript before starting a new chat.
2.Include project/model/scenario IDs, selected component, ingest receipts, result IDs and the accepted limitation text.
3.Remove credentials, account emails and private raw data before sharing.

Expected evidence

✓ A text packet that another authorised operator can replay in the same organisation.

Trust check

Do not embed API keys, organisation secrets, signed URLs or restricted raw rows in prompts or public documents.

Before you share

Completion checklist

Active model, scenario and selection are visible.
Prompt names target, value, unit, action and verification.
Every mutation appears in a typed action chip.
Ambiguous names are resolved, not guessed.
Workflow inputs and result IDs are recorded.
Observed, derived, assumed and modeled claims are separated.
Transcript handoff excludes credentials and restricted data.

Answer desk

Frequently asked questions

Can the Modeler Agent edit node and asset parameters?

Yes. It can edit validated fields on selected or unambiguously named components through scenario overrides, then read back the saved values.

What happens if two components have similar names?

The Agent should refuse the ambiguous partial match and ask for the exact name or a UI selection. It should not guess.

Can an Agent run a BESS or Dholera study?

Yes. Modeler exposes typed first-party workflow tools. Review their arguments and the returned result IDs because these tools make model changes and launch solves.

How should an agent cite a Modeler result?

Include project/model/scenario, result ID, created time, solver/formulation, modeled horizon, snapshot stride, source/ingest receipts and the exact truth label.

← Previous150 MW data-centre example