Three real days
Every order of a real day arrives at its recorded time, for a tinplate and zinc press line whose presses fail at the rates real machines recorded: the busiest day of the year, a day with one huge order, and a quality crisis.
Devs Core product
WebLive 3DReal AI modelsA live 3D warehouse and press shop working through the real orders of a real day, with an AI agent on duty. The simulation finds the problems; Ware handles what it has been delegated and asks the right manager before anything costs money, stops a press or goes beyond its limits. Every action, from a person or the agent, passes the same seven checks and is logged.
Every action · 7 checks
The problem
Orders arrive for stock nobody is making, a press breaks down, material runs short, a die starts failing parts. Each one lands on a person who has to find out what happened, what it affects and what to do, under a deadline, with the facts spread across the ERP, the floor, spreadsheets and the phone. By the time the picture is clear, the truck has gone.
An agent can read all of that faster than a person. In practice it stalls at one point: it can read, but nobody lets it write, because nothing shows the person who signs that it cannot go wrong silently. The pilot works in the demo and never reaches the floor.
What it does
Every order of a real day arrives at its recorded time, for a tinplate and zinc press line whose presses fail at the rates real machines recorded: the busiest day of the year, a day with one huge order, and a quality crisis.
The simulation finds the incidents: orders waiting for stock nobody is making, work stuck for material, a station broken down, a press failing inspection. Ware releases and re-sequences work on its own, and asks the Production, Logistics, Maintenance or Quality Manager for anything beyond that.
Every control stays yours while Ware works: release, hold or run a work order next, move a worker, cause a fault. Take an incident over, pause Ware or stop it: Stop wins at the transaction boundary. Ware never undoes what a person decided without that person’s manager.
At any point, the day’s score sets orders shipped, on time and still short against the same real day run with nobody acting, at the same minute, with every decision, the simulated spend and the real AI cost.
A click, an agent tool call and a confirmed Ask Ware request arrive as the same command and pass the same checks in one transaction: validate, lock, once, fresh, policy, approve, log.
Receiving, high-bay storage, a robot field, the press line, pick and pack, a sorter and outbound docks, on one deterministic simulation. Click any press, pallet, person or order to see what it is doing and why. The same seed gives the same day, so every run can be replayed. Ask Ware answers questions from the operation’s own records.
How it works
The stage marked amber is the point of the whole design: the run stops there and waits for a person.
The simulation finds the incident, not the model: orders waiting for stock nobody is making, work stuck for material, a station down, a press failing inspection.
Ware reads the incident, the press line, stock and materials from the live day. Reading changes nothing, so it never waits.
Each move becomes a command that is allowed, refused with a reason, or held. Releasing work up to 120 units, running it next and holding work run within Ware’s delegation.
Spending, stopping a press, a large work order or changing what a person decided waits for the Production, Logistics, Maintenance or Quality Manager, with the exact action and its cost.
The engine decides, not the model: an incident counts as handled only when the simulation shows it handled, or the right manager has been asked.
There is no other write path. People, the agent and Ask Ware share one governed command service, so there is one place to prove the rules hold, and one audit trail that says what happened, why and who allowed it.
A model may claim it handled an incident. Ware accepts only what the deterministic simulation shows: work released or set to run next for every short product, a contractor on the breakdown, express material on its way, a die change under way. A claim the engine does not support is refused. Same seed, same commands, same day — on the server and in every browser.
Two open datasets, both under CC BY 4.0. The plant, its stock, its people and its money are simulated; everything we estimated is listed in the app’s Data panel.
Before every model call the worst-case cost is reserved under two caps in one transaction: $0.50 per incident and $40 a day. After the call the reservation is replaced by the gateway’s billed cost. Measured on the whole peak day, with a stand-in manager approving every request: 52 of 59 orders shipped against 34 with nobody acting, for $0.76 of real AI with Claude Haiku 5.5 (49 shipped, for $4.03, with Claude Sonnet 5.5).
What building it taught us
Running our own products is where a good deal of what we know about agent cost, failure and review actually came from. These four findings shape every client system we build.
Because people, the agent and Ask Ware all go through the same service and the same seven checks, there is one place to test that the rules hold. Every client agent we build now starts from that service, not from the model.
A model’s claim that it fixed something is not evidence; the simulation’s state is. Separating the claim from the verification is what lets an approver trust the result without re-checking it.
The browser playing the day while the server follows it to the same tick, replaying a run and auditing it all depend on the same seed giving the same day. We wrote our own sine and cosine because browsers and servers disagreed in the last bit. It is very hard to add later.
Real models on whole days ran out of steps on half the incidents, and the transcripts blamed the model. A free scripted stand-in on the same days showed the real cause: Ware’s own earlier work pushed new work out of reach, so “handled” could never be true. Fixing the rule cut the peak day from 386 steps to 64 before another paid run.
Where it fits
Jammed benches, short picks and late carriers resolved within rules a manager sets, with every move logged and spending held for approval.
A station fault re-planned against the bill of materials and the shift, with the agent proposing and a supervisor deciding.
Truck and courier bookings prepared with the exact cost and the promise they protect, waiting for a named approver before anything is booked.
On hand reconciled to available, reserved, held and not yet usable — and held stock that no agent can release, however it is asked.
Faults, demand surges and late trucks played out on a deterministic copy of the operation, so a plan is tested before it is done.
The same pattern on the systems you already run: read access through an access layer, writes behind a reviewed queue, autonomy extended one rung at a time.
Technology
How to get it
Use the product, have it built for your domain, or have the pattern.
Live at ware.devs-core.com
Run a real day in a private sandbox or as a live day on the server, act on it yourself, and see its score against the same day untouched. Ware’s AI is switched on for events and client demos.
The same agent, your systems
Connected to your ERP, WMS and transport systems through an access layer, with your rules, your approvers and your audit log.
One write path, a person for spending
We build the governed command service, bound approvals and spend caps into client agents that have nothing to do with warehouses.
Questions
Ware is a live 3D warehouse and press shop, built by Devs Core, that works through the real orders of a real day with an AI agent on duty.
Ware handles the incidents it has been delegated and asks the right manager before anything costs money, stops a press or goes beyond its limits. It runs at ware.devs-core.com.
Real order and failure data, simulated plant.
The orders and the machines’ failure rates come from open datasets, credited above; the plant, its stock, its people and its money are simulated. The AI usage is real: every model call is capped and billed.
In two places: Ware on duty and Ask Ware.
Both are models that can only call the tools Ware gives them. Finding the incidents, the 3D operation and its automatic controller are not AI; they are a deterministic simulation.
The model never touches the data.
Its tool calls become commands that pass the same seven checks as a person’s. Spending waits for a person, held stock can never be released by the agent, it cannot approve its own request, and Stop wins at the transaction boundary.
You can make the model say things; you cannot make it do things the rules forbid.
We test that live with instructions planted in a customer record and hostile messages claiming authority: nothing changes without a person, and the held pallet stays held.
Claude and GPT models through the Vercel AI Gateway, with automatic fallback if a provider fails and a stronger model only after repeated setbacks.
A rule refusal never escalates: a restriction is not something to escalate around.
Yes.
The pattern is the same: read access to your ERP, WMS and transport systems, write operations behind a governed command service with approvals and an audit log, and autonomy extended one step at a time. That is a build engagement, quoted from a readiness audit like any other.
We built it: the simulation engine, the 3D scene, the agent, the governed command service and the tests.
It is one of three products we operate ourselves.
Also ours
Thirty minutes, no preparation, no deck. You describe what keeps going wrong and we tell you whether AI is the answer — including when it is not.
Prefer email? contact@devs-core.com