Research project

Reality Pack – makes physical energy assets securely manageable by AI agents.

When an AI agent controls energy assets – charging and discharging a battery, regulating what goes to the grid, switching an operating mode – the quality of the model is not what makes it safe. What is needed is an independent system for permissions and limits, and that is what we are building with Reality Pack, a research project. An agent has to know which actions it may take on its own, which need a human, and which are forbidden. We turn technical limits, regulatory requirements, contractual agreements and operator instructions into machine-readable rules, and a decision made against them can be traced back to the rule and the source behind it. We start with battery energy storage.

Get involved
The challenge
Software already decides well what an energy asset should do. Whether it may do it is a different matter. Technical limits, grid connection terms, contracts and mandates are scattered across manufacturer, grid operator, marketer and owner, and they are written for people. Every system has to work them out again, and when it gets one wrong, the operator is liable.
Our approach
We describe a real asset once, in machine-readable form, across five layers. Before it acts, any control system, an AI agent included, asks and gets a deterministic answer: permitted, refused or unknown. The answer names the limit that applies and a source a human can check. We do not replace BMS, EMS, optimiser or marketer. We sit in front of the action.
Who it is for
Organisations responsible for many professional assets, where responsibility, control and knowledge of the rules sit with different parties: operators of large storage, technical operations managers, asset managers, direct marketers and aggregators. Makers of AI agents are not the first customers.

Research project, starting with one battery. No customer and no price yet. What runs today →

A stand-in, not the target. Reality Pack is aimed at large battery storage run by operators, operations managers, direct marketers and aggregators (B2B). We start on a small device because every limit on it can be checked cheaply: the Anker SOLIX Solarbank 4 Pro, one unit, two modules with 1,000 Wp, never discharged below 20 %. The day shown is simulated, not measured. The device has been ordered and the gate has not run against it yet.

The solution

The agent acts, the pack bounds

A battery inside five contours, an agent outside A battery sits inside five contours of its own shape, one per layer: physical, digital, normative, economic, institutional. Two are drawn lighter because those layers are not written yet. A single channel leads out to the right, and each contour ends at a threshold inside it. An agent stands outside. One impulse from the agent passes all five thresholds and reaches the battery; a second is stopped at the normative threshold. PHYSICAL · DESCRIBED DIGITAL · TO BE PROCURED NORMATIVE · DESCRIBED ECONOMIC · DESCRIBED INSTITUTIONAL · ROADMAP refused · normative battery

AI agent

“Maximise today’s economic return from this asset.”

50 MWh of storage. To deliver that, the agent has to dispatch the asset – charge, discharge, the physical layer – and sell its flexibility on day‑ahead, intraday, FCR and aFRR, the economic layer.

The institutional layer then asks what neither of them answers: may it offer 20 MW? Sign a binding contract? Dispatch the battery? Put €50,000 at risk?

Five layers, one question

The customer’s agent asks the pack before every action: is this action, for this asset, right now, executable, permitted and sound in the Bilanzkreis? Unlike an optimiser, the pack does not replace the trading decision – it is what makes it permissible. Layer 4 is not written and stands here so the gap cannot be overlooked.

Layer 0 Data connection Do we know — and how old is what we know? procured
  • Access path, protocol, update rate
  • Age and latency of every value
  • Other parties holding a write path
  • Telling a refusal from an unreachable source
  • Retention and audit trail

Numbered 0 because it is not a test of its own: the other four ask whether an action is possible, permitted, worthwhile and mandated — this one asks whether we know anything at all, and how old it is. Not an assumption but a procurement: with an open local interface the access belongs to the operator, otherwise to someone else’s cloud, which can withdraw it.

SourceEMS/BMS exports, schedule data, market data — via exports in a pilot, live in operation

Layer 1 Asset physics Can the asset do it right now? included
  • State of charge and power history
  • Available charge and discharge power
  • C-rate, round-trip efficiency, temperature window
  • Degradation budget
  • Reserve held for balancing markets

SourceData sheet, EMS/BMS, operator settings

Layer 2 Grid and regulation Is it allowed to? included
  • Grid connection agreement, connection capacity
  • Prequalification status per balancing product
  • Running and announced redispatch calls
  • Grid fees, § 118 (6) EnWG
  • Reporting obligations

SourceGrid operator, prequalification notices, rulebooks

Layer 3 Market and Bilanzkreis Is it worth it, and what does deviating cost? included
  • Submitted schedules (Fahrpläne)
  • Capacity committed by balancing-market awards
  • Open positions
  • Imbalance prices (Ausgleichsenergie)
  • Gate closure times, nomination status

SourceTrading systems, Bilanzkreis settlement, exchange data

Layer 4 Mandate and delegation May THIS agent do it, for THIS principal? roadmap
  • Who may act for whom
  • Up to which limit, with what accountability
  • Where the mandate ends and a human decides

Not written. The pack answers whether an action is permitted; it does not yet answer whether this agent may carry it out for this principal. That is exactly where the step from shadow operation to live operation fails today — so this is the next stage, not a footnote. Milestone 05.

SourceMandate of the owner, delegation rules, accountability chain

What a No looks like

The pack is meant to answer Yes, No with a reason, or escalation to a human – every answer logged and traceable to the layer that produced it.

The five verdicts below are hand-written examples – no running system produced them. They show the shape a refusal needs for a human to check it.

  • No An aFRR call is running, 5 MW are committed until 14:15. Market and Bilanzkreis
  • No The schedule for the 13:00 quarter hour has been submitted. A deviation is settled as imbalance energy. Market and Bilanzkreis
  • No A redispatch request for 14:00 to 16:00 is on file, capacity is reserved. Grid and regulation
  • Limited SoC at 92 %, charging power down to 3 MW instead of 10 MW. Asset physics
  • Escalation The FCR prequalification lapses in 30 days, the bid is not released. Grid and regulation

Whoever owns an asset can let any agent near it. Whoever builds agents can reach any asset. Direct marketing becomes something software can do, instead of something that takes a trading house.

The hypothesis

What we claim —
and how it could be wrong

What is missing

Computing what to do with a battery is solved and sold. Computing whether an action may be executed is not. The relevant constraints are distributed across manufacturers, grid operators, regulators, markets, contracts and asset owners – written for humans and often dependent on context. Today, every agent or application has to reconstruct this knowledge for itself.

The hypothesis

These constraints can become a machine-readable property of the asset rather than knowledge embedded in each agent. If a real asset is described across five layers – digital, physical, normative, economic and institutional – a foreign agent should be able to propose an action and receive a deterministic decision on whether it may be executed, together with the applicable constraint and a source a human can verify. One maintained description per asset would allow different agents to operate against the same boundaries without having to know them in advance. The Battery Reality Pack tests this hypothesis for a grid-scale battery; the same principle could later apply to PV plants, wind turbines, data centres or charging infrastructure.

What would refute it

The hypothesis fails if relevant constraints cannot be represented without modelling their entire surrounding context; if every asset requires a bespoke, hand-written description; if a refusal cannot be traced to a verifiable source; if an agent needs prior knowledge of the schema or the asset type to use it correctly; or if foreign agents never arrive at all – if operators keep running their own assets with their own software, this is a schema nobody reads. The decisive observable is therefore simple: a foreign agent proposes an impermissible action, the system refuses it correctly, and the refusal can be independently verified. Until that happens, this remains a hypothesis.

Why we do this

Portrait of Dr. A. Martens

Dr. A. Martens

LinkedIn ↗

AI agents are becoming capable of making decisions and taking actions autonomously, but operating a physical energy asset requires them to respect technical limits, regulatory requirements, contractual obligations and economic constraints. Today, this knowledge is fragmented across systems, documents and organizations and has to be reconstructed for every new application.

The motivation behind the Battery Reality Pack is to make these boundaries an explicit, machine-readable property of the asset itself, enabling AI agents to operate real energy assets autonomously while reliably staying within what is permitted.

Once a month, by mail

The logbook, sent

What moved, what failed, where the count of reasoned refusals stands. At most one e-mail a month, and none in a month where nothing happened. No product news, because there is no product.

You confirm by clicking a link we send you. Until you do, nothing is stored anywhere; after you do, your address sits in one mailbox and in no other system. One e-mail to maw@marts.group takes it off again.