PromptFactory · by PromptCompany
A product engineering factory run by AI agents: from task to tested code, with a human deciding.
Like ChatGPT, but instead of a conversation you get engineering work: tested code and circuit boards designed in code. Mechanical CAD is next. A human sets the rules and makes the call at the checkpoints.
What we do
Software and electronics, built by AI agents. A human sets the tasks.
For teams building software or electronics products. Each task goes from your backlog to code merged on main through fixed stages: build with tests, review and quality gates. Larger tasks also get a proof stage, and the largest a plan first.
- Taskfrom the backlog You set tasks and priorities
- PlanTHINKlargest tasks
- BuildBUILDevery task
- Test and provePROVElarger tasks red tests → another build round
- ReviewREVIEWevery task review findings → another build round a second, independent review for larger code changes and every risky change
- Gatesquality When a gate asks, a human decides
- Merge to mainLANDING one landing queue, in order
In numbers
Read from the factory's ledger and the git repositories.
- 1,203 tasks merged on main since 22 Sep 2026 A task is one change with its tests, in product code or in the factory itself. as of 3 Nov 2026
- 72% of tasks since 15 Oct had two independent reviews 469 of 655 tasks merged on main. as of 3 Nov 2026
- 56,411 ledger records since 6 Sep 2026 Every hash in the chain recomputed: no break, no mismatch. as of 3 Nov 2026
- 52.4 tasks a day in the best week, 19–25 Oct as of 3 Nov 2026
- 83 git repositories in 12 layers From firmware and contracts to backend, web apps and CI/CD. as of 3 Nov 2026
- 20 printed circuit board designs As code, in atopile and KiCad. as of 3 Nov 2026
Services
Software and electronics today
Expanding into: mechanical engineering with CAD as code in build123d, systems engineering, cybersecurity and manufacturing.
Software
Embedded firmware, on-board software (ROS 2), backend services, web apps, AI models, simulation, CI/CD.
Electronics
Printed circuit boards as code, in atopile and KiCad, with the BOM generated from the design.
AI agents design the boards in sessions with an engineer.
Technologies the factory works in
- C/C++ · Zephyr RTOS
- Python
- TypeScript · React · Next.js
- Qt/QML
- ROS 2
- Gazebo (SITL)
- PyTorch · ONNX
- FastAPI · gRPC · PostgreSQL
- Protobuf
- Docker · GCP · CI
- atopile · KiCad
- BOM
Reference project
SmartDrones sp. z o.o. (a related company). We proved the factory at home first. The SmartDrones team uses it every day to build its own drone-inspection software: firmware, on-board software, cloud and apps. The same team also designs the drone's printed circuit boards as code, with AI agents in sessions with an engineer, and tests the firmware in SITL simulation. The code, the designs and the results belong to SmartDrones.
The work spans:
- Autopilot and navigationembedded software, C/C++, Zephyr RTOS
- Mission computer and ground stationsROS 2
- SITL simulationflight in the Gazebo simulator
- AI perceptionwind turbine blade inspection; PyTorch, ONNX
- Backend servicesFastAPI, gRPC, PostgreSQL
- API contracts and SDKsProtobuf; SDKs generated from the contracts
- Web and desktop applicationsTypeScript, React, Next.js; Qt/QML
- CI/CD and end-to-end testsPython; Docker, GCP, Robot Framework
- Electronicsthe autopilot family and modules: sensors, power, networking
Pilot
Run a pilot on your own project
The core of the factory is running today. We are looking for partners to prove it on their own project.
-
1 · ABOUT 2 WEEKS
Onboarding
A front-engineering workshop on requirements, backlog and architecture, then a copy of the factory set up for your project.
-
2 · 2–4 WEEKS
Factory run
The factory delivers tasks from your backlog.
-
3 · AT THE END
Results
A review of the features delivered, the measurements, and a pilot report.
WHAT YOU BRING
- a project or repository
- one person on the product side
- about 2 hours a week
WHAT YOU GET
- the features delivered in the pilot, merged on main
- the measurements from the run
- a pilot report
This is the proposed shape of a pilot; we agree the details on a call.
How we work
From backlog to main, by the same rules every time
A task is one change with its tests, and its size sets the stages it goes through. AI agents work in roles; a human sets the tasks and the rules, and decides wherever a gate asks.
The size rule
| Task size | PlanTHINK | BuildBUILD | ProofPROVE | ReviewsREVIEW |
|---|---|---|---|---|
| XS, S | no | yes | no | 1 |
| Larger | no | yes | yes | 2 for code |
| Largest | yes | yes | yes | 2 for code |
| Risky change, any size | stages by size | 2 | ||
2 means a second, independent review: larger code changes and every risky change get one. Red tests and review findings send the change back for another build round. Every task ends at the quality gates and the one landing queue.
Today the reviews run on Claude models (Anthropic). We plan for the next versions of the factory to be model-agnostic.
Owner's console and app
Questions arrive as cards. A decision is one tap.
When the line needs a human, the question shows up in the console and in the app: a React and TypeScript web app, and an Android app with a microphone. Its tabs are Conversation, Decisions and Health, and you can talk to the factory by voice.
- you set the tasks and their priorities
- when a gate asks, a human decides
- for risky changes, the owner can also veto
-
Hash-chained ledger
Every step is a record in an append-only ledger. Each record carries the previous record's hash and its own SHA-256 over canonical JSON, so an edited or deleted record breaks the chain.
-
One legal next step
The line advances by reading the ledger. From any state there is exactly one legal next transition.
-
One landing queue
Every waiting merge holds a ticket in one line. Only the first in line tries the lock on its repository, so merges reach main in order.
-
Quality gates
Checks a change must pass before it merges. When a gate asks, a human decides.
-
Budgets and round limits
Each task and each agent has a token budget; past it, a task can be held for the owner's decision. The number of build rounds is capped by task size.
-
Machines on demand
A build or test job gets a machine started for that job; a cache hit needs none. Simulated flights run on their own VM, one per flight, deleted when the flight ends.
-
Agents in roles
Planning, build, test and review agents work the line. A conductor and management seats sit above it.
-
Front engineering
Before the factory starts building, we set the requirements, write the backlog and the architecture together with you.
Testing
Tests matched to each change, the full battery every night
A build runs the task's acceptance criteria and the test suites its change touches. The proof stage reruns exactly that scope. Every night the full battery runs.
Unit
functions and modules in isolation
pytest · GoogleTest · Vitest · Jest
Functional
each task's acceptance criteria
pytest · Playwright
Integration
services together, in containers
Docker Compose · pytest
End-to-end
full scenarios on a running system
Robot Framework · Playwright
SITL
software-in-the-loop flight in the Gazebo simulator
one VM per flight, started on demand
In the repositories: 1,736 pytest files, 179 C++ files with GoogleTest tests, 67 Robot Framework files. as of 3 Nov 2026
When
- Every task
- the change's tests in the build, run again in the proof stage when the task has one
- Every night
- the full battery
Agents test hands-on
Playwright MCP: an agent drives a web app in the browser.
Robot Framework MCP: an agent runs and reads end-to-end tests.
About the company
PromptCompany, the company behind PromptFactory
PromptFactory by PromptCompany is a product engineering factory run by AI agents. The core of the factory is running today. We are looking for partners to prove it on their own project.
- PromptFactoryour product: a product engineering factory run by AI agents
- Software and electronicswhat the factory builds today
- Lublin, Polandheadquarters
Get in touch
Let's talk about a pilot on your project.
Tell us what you are building and we will set up a short call about a pilot.
- Email: office@promptcompany.ai
- Phone: +48 785 085 505
- Address: PromptCompany sp. z o.o.
ul. Frezerów 2
20-209 Lublin, Poland