Project: How far can we go in building an energy company operated by AI Agents?

Logbook

1. 14 September 2026 Decision Layer 4 is not trivial at 5 kWh — and I said the opposite four days ago

   Two layers of the five are not written: the digital one and the institutional one. Until today the page labelled them “the two layers the pack does not contain”, which reads as out of scope. It is not what was meant, and it contradicted the paragraph below it, which calls Layer 4 “the next stage, not a footnote”. The label now says still to be written.

   Both are in scope. They are in different states, and the difference matters.

   The digital layer is being bought right now

   It was marked presupposed until yesterday because the original version addressed a direct marketer, who already has EMS access, schedule data and market data. For a private operator that inverts. The data connection is not a precondition, it is the first thing that becomes real — before the schema, before any runtime.

   What is still missing is the description: where the data comes from, at what rate, with what latency, who else holds a write path, and how an agent tells a refusal from an unreachable source. That last one is not a detail. From the outside the two look identical: nothing happens.

   The institutional layer: I called it trivial, and that was wrong

   On 2026-09-10 this log said the mandate question was trivial for a private setup — the battery belongs to the household, so there is only one principal. That is wrong. There are at least five, and every one of them is real at 5 kWh:

   1. The manufacturer’s own optimiser. It evaluates weather, consumption and exchange prices by itself. Two agents, one asset. Which one wins when they disagree?
   2. The warranty. Ten years or ten thousand cycles. An agent that cycles hard spends a promise that was made to someone else.
   3. The household. If a fridge hangs off the backup socket, “discharge to zero because the price is good” is a decision about somebody who was never asked.
   4. The supply contract. A dynamic tariff has terms.
   5. The law. The operator entered in the Marktstammdatenregister is a person. The agent has no legal personality; there is nobody else to be liable.

   The structure is identical to the 20 MW case. Only the amounts are smaller: an agent acts on an asset, and the consequences land on a person who is not in the loop.

   The sharpest instance costs about €1,600

   A consumer device keeps a write path for its manufacturer that the owner cannot cancel without losing functions.

   The owner is not the only principal, and cannot revoke the other one.

   On a grid-scale asset the same question would be settled in a contract. Here it stands bare. For testing the idea, that makes this the better case, not the weaker one.

   What follows

   Milestone 05 is declared: a refusal that comes neither from physics, nor from law, nor from price — the action is executable, permitted and worthwhile, and the agent still may not carry it out.

   It is deliberately declared without a target date. Every other milestone has a window; this one has a predecessor instead. A date here would be a guess dressed as a plan.

   Nothing is built. The status above is unchanged.

2. 14 September 2026 Decision An address: reality-pack.dev — the series, not the instance

   The project gets its own address: reality-pack.dev. Not battery-reality-pack.energy, and not a subdomain of an existing one.

   Why not .energy

   It would have contradicted a sentence in the first screen of this page: “It is the first case, not the goal.” An address that commits to energy says the opposite of that. Energy is where the layers happen to be hardest and best documented — it is not what the schema is about.

   It would also have named Pack #001 only. A PV plant, a wind turbine, a data centre and a charging park are already named in the repository as conceivable further packs; each would have needed its own domain. And the renewal price is roughly five times that of the alternative, every year, for a project with no revenue that is meant to stay readable after the twelve months are over.

   Why not a subdomain of an existing site

   It would have been free and, today, perfectly honest. The argument against it is the measure of this project: a foreign agent. A schema that lives under a company’s domain reads as that company’s schema. For something meant to be consumed by people who were not involved in writing it, a neutral address is a small but real signal.

   Registering is not publishing

   The obvious objection to buying an address for a schema that does not exist is that it is an advance on something unbuilt. It is not, and the distinction is worth keeping: the build rule binds what is delivered, not what is reserved. An unused domain makes no claim to anyone. A page that asserts things it cannot hold does.

   What is deliberately not done yet

   The site configuration still carries no site value, so the delivered pages emit neither a canonical URL nor an og:url. Checked today: the domain has no DNS delegation and answers nothing. Pointing metadata at an address that does not respond would be exactly the kind of claim this project keeps removing from its own files.

   That one line gets set when the domain actually answers. Deployment itself is a separate question and not decided here.

   The shape, once there is something to deploy

   While there is only one pack, the page stays at the root. A /battery path gets created when there is a second pack for it to be distinguished from — an index page above a single entry is structure without content.

   The status above is unchanged. An address is not a schema.

3. 13 September 2026 Failure Three errors in our own files — the build rule applied to everyone except us

   Going through the layers turned up three errors. All three sat in our own files, all three were found by reading, none ever produced an error message.

   1. The README promised five layers, the page held three

   The README said: “a machine-readable representation of a real asset across five layers”, followed by a table with five rows. The page has always said: Layer 0 presupposed, Layer 4 roadmap, three are included.

   Both texts were written on the same day. Nobody lied — the sentence had two jobs and did only one: it defines what a Reality Pack is in general, and directly below it says that this one is Pack #001. A reader concludes five.

   That is exactly the promise the build rule forbids — stated in the same file, two sections further down. A promise may only say what the artefact holds. Corrected: the table now has a third column, and the sentence before it says three.

   2. “Layer 0: presupposed” was wrong, and at the most expensive point

   The digital layer was marked presupposed. Choosing a real device showed it is nothing of the sort: it is decided by the purchase. The same manufacturer’s range offers devices with an open local interface and devices without. With the latter the access belongs to someone else’s cloud, which can throttle or withdraw it.

   And the consequence the word “presupposed” conceals: an agent that cannot tell a refusal from an outage does not have this layer. From outside the two look identical — nothing happens.

   The state is now called procured. Not a nicer word, a truer one.

   3. The institutional layer is Layer 4, not Layer 5

   An entry dated 2026-09-10 said “Layer 5” twice. The stack counts from zero: 0 digital, 1 physical, 2 normative, 3 economic, 4 institutional. There is no fifth. Corrected.

   What the actual finding is

   Three errors, found by laying our own files side by side. None of them would have surfaced through a build, a test or a checksum, because none of them is a technical defect. All three are statements that had stopped being true.

   That is the same class as the button which pointed at a deleted section two days ago, and the same as the duplicated status display in the footer: rules only bind going forward. Backwards, someone has to look. The build rule was in the README from the start — it did not prevent the promise two paragraphs above it.

   The status above stays unchanged. A corrected sentence is not progress, it is the withdrawal of an advance.

4. 10 September 2026 Decision Panels are included — bought for the uncertainty, not for the electricity

   The test object includes solar panels, up to 2,000 Wp — the ceiling of the simplified rules. The inverter is already inside the battery, which brings four MPPT trackers and accepts up to 5,000 Wp of solar input. So what is missing is panels and a mount, nothing else.

   They are not part of this experiment because they generate electricity. They are part of it because without them two layers stay empty.

   Without panels, Layer 1 has nothing to estimate

   A battery on its own is fully known: state of charge, power limits, temperature — all measured, nothing forecast. There would be no uncertainty in the interface that could be declared.

   That would remove the sentence the whole undertaking hangs on: an agent does not need the answer, it needs the declared boundary of the answer. With panels a quantity enters that the agent cannot command and has to predict. Only against that does it show whether the schema is more than a data sheet in JSON.

   Without panels, Layer 2 loses almost everything

   The battery alone would still be subject to registration: every stationary battery storage system must be entered in the Marktstammdatenregister — regardless of size, commissioning date and actual feed-in. So Layer 2 would not be zero.

   But everything concrete hangs on generation: the 800 W feed-in limit, the 2,000 Wp module ceiling, the aggregation per metering point. That is precisely the rule set on which the idea of two devices failed — this project’s best evidence so far. Without panels it would not be there.

   An open finding that belongs on the record

   Whether a battery that charges exclusively from the grid and never feeds in fits cleanly into that rule set could not be settled. German renewables law defines a plug-in solar device as modules, inverter, connection cable and plug — a battery does not appear in the definition.

   That is itself a Layer 2 finding: a real, purchasable configuration for which the rule has no clear category. Not “forbidden”, not “permitted”, but not provided for. An agent that depends on an unambiguous answer has none here — and must be able to say so instead of picking one.

   What Layer 3 gains

   Configuration	The question put to the agent
   Battery only, dynamic tariff	Is electricity cheap right now? One source, one decision.
   With panels	Do I take the cheap hour now — or keep capacity free for tomorrow’s sun?

   Only then do two uncertain sources compete for the same storage. That is the question stated further down this page: which permissible action is worth most?

   Open

   • Location and orientation are not decided. Without a usable spot this stays theoretical.
   • Whether panels arrive with the battery or later is open. The battery and a dynamic tariff are enough for the first reasoned refusal; the panels widen the test, they do not enable it.

   The status above stays unchanged. Still nothing bought, nothing built and nothing refused.

5. 10 September 2026 Decision The test object will be a home battery — and the first real refusal came while shopping

   Decided: the test object will be an Anker SOLIX Solarbank 4 E5000 Pro (5 kWh, ~€1,600), together with a dynamic electricity tariff. It has not been bought, so milestone 1 stays open.

   Why this device

   It has what matters for a foreign program: an official local Modbus TCP interface, read and write, enabled in the manufacturer’s app. No cloud detour, no reverse engineering. For home batteries, local write access is the exception, not the rule.

   The dynamic tariff is not an add-on, it is the reason. It makes Layer 3 real: hourly exchange prices are a limit nobody in this house set. A refusal that goes back to a rule we wrote ourselves proves nothing.

   What is not real about this asset belongs here just as much: Layer 2 is simulated. The device feeds in at 800 W and is therefore far below the 4.2 kW at which § 14a EnWG applies — it will never receive a control command from a grid operator. The rules it has to be measured against are ones we write ourselves. That is a weakness of the test object, not of the schema, and it stands here so that nobody later mistakes it for a success.

   The first real refusal came before the purchase

   The idea was to stack two devices to get past 4.2 kW and make Layer 2 bite. It failed on three independent grounds, and the failure is the best evidence so far for this project’s thesis:

   1. Expansion adds kWh, not kW. Up to five extra modules take it to 30 kWh. The inverter stays at 2,500 W.
   2. The 2,500 W is not the grid figure at all. The manufacturer writes “800 W grid-coupled” and “2,500 W off-grid output” — the larger number belongs to a backup socket, not to the grid. With an electrician and a Wieland connector, 2,500 W of feed-in is possible; but then it is no longer a plug-in device.
   3. Two devices on one metering point are less permitted, not more. The 800 W limit applies per metering point, not per device. Two devices make 1,600 VA and lose the simplified regime entirely.

   An agent that reads only the data sheet plans 2,500 W of feed-in and is wrong. An agent that adds the second box arrives at 5,000 W and is wrong not by a factor but categorically: the result is not a larger number, it is a different legal regime. The same hardware may do different things depending on how it is installed and registered. A schema that describes only the device cannot answer that question in principle.

   Two observations for the schema design fell out along the way:

   • Power limits are direction-dependent. Charging from the grid up to 2,500 W, feeding in 800 W, island output 2,500 W. Three numbers, three rules, one device. A single field “maximum power” would already be wrong.
   • There is already an optimiser on the device. The manufacturer evaluates weather, consumption and exchange prices itself. Who holds the mandate when a foreign agent wants something else? That is Layer 4, unplanned and free.

   Possible later stage: a sonnenBatterie in a virtual power plant

   Noted, not decided. It is the only asset at household scale I know of where Layers 2 and 3 come from outside:

   • Real prequalification for frequency containment reserve — sonnen’s virtual power plant is prequalified with TenneT (2018), Amprion (2021) and TransnetBW (2024), marketed through Next Kraftwerke.
   • Real committed capacity: while an FCR call runs, the battery is obliged. That is literally the example further down this page — “No, a call is running, capacity is committed until 14:15” — only small and real.
   • A real conflict of mandate: the household owns the battery, the pool dispatches it. Layer 4 stops being trivial.

   The price has to be named honestly: roughly €8,000 to €12,000 installed instead of €2,000, against a profit share of up to €100 a year. Financially that is not a business. It would be an expenditure for insight, and it only pays once the schema carries at the small asset at all. Before that it would be an advance.

   Open

   • Not bought. Milestone 1 stays open until a real asset is named — with a data source and permission to use its data.
   • Has anyone written over Modbus to this exact model? One source says a sibling model is the hardware-validated one and the Solarbank 4 merely shares its register map. That is the difference between “tested” and “should fit”.
   • The 800 W limit is to be settled at the source, not at a blog article. While reading up, sources contradicted each other: one headline said “abolished”, the body text of the same page said “unchanged”. That is exactly what Layer 2 is for.

   The status above is not changed by this entry: no schema, no runtime, no users, zero reasoned refusals. A purchase decision is not a refusal.

6. 10 September 2026 Open Day 0 — there is nothing

   The experiment starts today, and today none of it exists: no schema, no runtime, no users. That is why the four values above stand where they stand.

   What does exist is a question, a frame of twelve months, and four milestones whose criteria are fixed today — before anything was measured, so that they cannot move later.

   What happened today: the technical description of the Battery Reality Pack was moved here from the Agent Reality Stack. It stood there as a product page. This project sells nothing, so the commercial layer has been removed — pilot, annual licence, operating contract, channels, the booking prompt. The technical content stands unchanged below it.

   Three promises stood on that page which it did not keep. Two are corrected: an unfilled placeholder for a monetary amount sat in the middle of the delivered text and is gone — not repeated here, so the string does not end up back in the text nobody wants it in. The “refusal demo” executes nothing and now says so itself: there are five hand-written examples. The third promise was that the page answered with 404 on the production host while being advertised. That is no longer a promise either: this page is deployed nowhere.

   The beginning of an experiment is a beginning. Anyone reading back in twelve months should be able to see what was not there on this day.