The digital twin for the agentic grid.

Distribution networks built for one-way flow now manage two-way flow, midday solar export, evening EV import, and clustered controllable loads. Top-down models are slow, expensive, and opaque; networks can rigorously model only a handful of the proposals they receive and triage the rest by judgement. Energy Queensland named the scenario-throughput bottleneck as the core constraint — AgenticGrid removes it.

AgenticGrid is a distribution-network digital twin that simulates, plans, and certifies — all from one model, all traceable to the rule that governs it.

A planner asks a question in natural language. A governed Claude Agent SDK skill connects to approved data, builds the location-specific model, runs the scenario, and explains the result — every step a logged tool call.

Take the Product Tour

20 slides of every screen.

National distribution-area map down to per-transformer voltage compliance and the AI-driven CBA. The fastest way to see what we mean.

Placeholder · 01 Distribution Areas

01 / 04

Distribution Areas

Australia’s entire distribution network, on demand.

AgenticGrid plots every Australian distribution area on one map — feeders, zone substations and modelled regions. Click any area, then ask the agent anything about it. The right pane shows AgenticGrid replying to a 150-word "what can you do?" prompt, in the VS Code chat.

Open full tour

One codebase ships two surfaces. Engineers and analysts get the VS Code Extension — AgenticGrid webviews sitting next to Claude Code chat under the ⚡ Synth Activity Bar. Utilities get a per-tenant Docker image on Azure Container Apps, branded URL, isolated data share, deployed in minutes by a single skill.

VS Code ExtensionHosted Web App

For engineers · runs locally

VS Code Extension

AgenticGrid lives as webviews under the ⚡ Synth Activity Bar, alongside the Modules launcher, scenario tree, and All Loops. Chat is VS Code’s own Claude Code panel.

• Local data + remote agents, in the editor you already use
• Open IDE chrome — drop into Python or Rust mid-question
• Same React, same skills, same backend as the web app

One React codebase, bundled by esbuild for both. Deploy a new tenant with the agenticgrid-deploy-webapp skill.

The same backend. The same skills. The same Used 4 tools: Skill, Bash, Bash, Read machinery. Only the system prompt and the voice differ. A TNSP planner, a regulator, and a 15-year-old all run the same simulator — but each gets a response shaped for them.

AgenticGrid is the first module on Synth: a modular agent platform combining the Claude Agent SDK, Ruflo orchestration, per-tenant deployment + skills, and structural memory. Build a new module, ship it as a per-tenant web app, share skills across tenants — all under one architecture.

12 scenarios × 61 sub-scenarios on a 9,556-premise, 300-substation network — all under current AER inputs. The numbers below are reproducible from the source data and code.

Finding 01

Solar is the binding constraint today

Broken Hill's network is already close to its solar-hosting limit. Push solar penetration to 80% and the breach is emphatic — hosting-capacity utilisation 178.9%, reverse power flow 40.3 MW (over 2.5× the peak import the network was built to carry). Unambiguous trigger for augmentation or active management.

Hosting utilisation

178.9%

Reverse power flow

40.3 MW

Baseline (as-built)

94.2%

Finding 02

The defensible non-network response is hybrid storage

Behind-the-meter batteries combined with community-scale storage emerge as the strongest economically viable option that stays within hosting limits — under current AER inputs (VCR $33.46/kWh, emissions $95/tCO₂-e, 5.5% discount, 15-year asset life). Beyond a point the binding limit is the network, not the economics of the batteries themselves.

Benefit-cost ratio

1.34

NPV (15-yr)

$9.08M

AER inputs

CBA Guidelines v3

Finding 03

A regulator-relevant negative result

Full-coverage two-day outage resilience via community storage does not pass an AER cost-benefit test — and full-coverage four-hour is little better at BCR 0.20. The lesson for non-network-solution planning is precise: resilience storage must stack benefits (demand management, emissions, energy arbitrage) and target genuinely critical load, not deployed as blanket coverage. Proving what doesn't work transparently is itself transferable knowledge.

Benefit-cost ratio

0.10

NPV (15-yr)

−$293M

Outage modelled

2-day full coverage

Each tenant ships with a system prompt tailored to its audience, runs on its own isolated Azure Files share, and reaches the same governed backend. Customer instances are sign-in-restricted; demo instances open with Microsoft sign-in (lead-capture coming soon).

Validation through the RACE for 2030 GridGuru Phase 2 program (Project 24.NT6.R.0928, Jul 2025 – May 2026), Industry Reference Group February 2026. Stated as organisational facts only.

Five audiences, one platform — each gets a tenant tuned to how they actually work.

PandaPower under the hood. AER CBA Guidelines v3 (Nov 2024). RIT-D framework. VCR $33.46/kWh (Dec 2024), emissions $95/tCO₂-e, 5.5% discount. Knowledge corpus across 8 organisations — AER, AEMO, AEMC, Essential Energy, Endeavour Energy, Ausgrid, Energy Queensland, ECMC.

Honest validation ceiling: the Broken Hill model reaches Pearson 0.6023 against measured transformer load (composite confidence 72.4% — "Validated"). The ceiling is transformer-level telemetry; a feeder-telemetry pilot is the documented next step. The platform reports its own confidence — it does not over-claim.