One prompt
becomes a factory.

You describe a production line in plain language. The Factory Agent plans the factory floor — a dimensioned 3D layout — balances the line, builds shift schedules, simulates throughput, writes digital work instructions, and programs the robot cells. Every program passes an 8-check kinematic verification before you see it marked verified.

Yes — with a catch that matters. The agent authors pick-and-place, palletizing, screwdriving, and welding programs on industry-standard 6-axis robot models with datasheet kinematics. But AI can get it wrong, which is why every program must pass joint-limit, continuity, timing, and collision checks (600+ samples per program) before it counts. Anything out of reach or over payload gets refused with the number.

Yes — that is the starting point. Describe the line in chat and the factory layout comes back dimensioned: stations, machines, conveyors, walls, safety zones, as a full 3D digital twin. If you have CAD, upload STEP assemblies and the agent extracts the BOM and places real part meshes into the station scenes. DXF floor plans and Excel data work too.

Both, in one loop. Shift schedules come from a deterministic CP-SAT solver — a real Gantt with due-date-risk flagging, not LLM guesses. Line balancing rolls cycle times up across stations, and throughput runs through discrete-event simulation: takt, utilization, and bottlenecks per station. A readiness gate tells you exactly what still blocks production, and why.

The code runs unmodified on the real robot controller software, and verification catches mistakes before they reach you — but your engineers stay in charge. An approvals workflow gates consequential changes, every edit records who made it (agent or human), and your PLC and safety engineers keep the final word.

Traditional OLP tools make you do the programming in a different interface. Here you prompt the task, the agent writes and verifies the program, compares multiple strategies, and optimizes cycle time — 27% faster on our benchmark cell. Planning, simulation, instructions, and robot programming share one context instead of living in separate tools.

About 30 minutes from prompt to first draft: layout, line balance, shift schedule, work instructions, and robot programs, with verification already run. From there you iterate by prompting — change the takt, swap a station, add a shift — instead of restarting a weeks-long planning cycle.

No. Emboscale layers on top of what you run today. It takes what you already have — DXF floor plans, STEP assemblies, Excel data — and hands back files your existing systems recognize: vendor-native robot code, STEP, DXF, XLSX, and AutomationML. Nothing gets ripped out.

Files your floor already recognizes: vendor-native robot code, deploy packs with safety-zone proposals and full verification provenance, STEP bundles, DXF plans, XLSX factory docs, and AutomationML.