For embedded, electronics and hardware teams

Code from AI agents that your team can check: a red test before the change, a green one after it, and the verdict of a separate review, recorded with the task.

AI agents write code faster than a team can check it. In firmware and hardware a defect is expensive, so what counts is whether you can show how a change was tested. A human sets the tasks and decides when a gate asks.

What stays with the task

  • The record

    A change package carries its cells: the test that is red before the change, green after it and red again when the change is reverted, plus mutations.

  • The review

    A separate review session reads the package: it rebuilds the change's tree and gives a SHIP or NO-SHIP verdict for that exact tree.

  • The gates

    By our rule no agent merges to main: a script lets the change in only after a SHIP verdict, and critical code waits for a human decision.

    By our rule, critical code needs two separate reviews and a human decision.

  • Coming soon: the same record in a Claude Code plugin that you install on your side.

How it works

A task comes from your backlog and goes through fixed stages.

  1. THINK: a plan before the build, for the largest tasks.

  2. BUILD: an agent writes the test and the code, and the change package carries the red and green cells.

  3. VERIFY: the tests in the change's scope run again, in a stage of their own for larger tasks.

  4. REVIEW: a separate review session gives SHIP or NO-SHIP, and NO-SHIP sends the change back to the build.

  5. Gate: the change reaches main in order, and when a gate asks, a human decides.

  6. Ledger: every step stays as a record in a hash-chained ledger.

In preparation: tasks straight from a product description in YAML, verification on separate machines, and one status page for the line.

Products

Claude Code plugins

Coming soon: pc-flow, a free Claude Code plugin under the Apache-2.0 licence, designed to hold a code change until the task's test has run and failed.

The plugin is designed to make no network calls of its own; Claude Code sends prompts and code to the model provider under your own account.

In preparation: domain plugins for software, electronics (PCB), mechanics (CAD) and simulation, each with one scope and a version.

Coming soon under the Apache-2.0 licence: pc-flow and the factory contracts, that is the profile schema and the gate contract.

For hardware teams

AI agents design printed circuit boards as code, in atopile and KiCad, in sessions with an engineer, and the BOM is generated from the design.

Before a board is released to the fab it passes the gates: connectivity, DRC against the fab's profile, availability of the BOM's parts, and the production package.

A separate review panel reads a board design change, and any doubt it reports goes back to a human.

Releasing a board to the fab and placing the order is a human act, never an agent's.

Our rule: board and part designs on a critical path do not go to outside AI design tools or cloud design services.

Coming soon: mechanical parts as code in build123d, with gates for mass, fit against the board and printability.

In preparation: the SBOM, a vulnerability scan and the product's threat model as records at release.

A paid pilot on your backlog

We run tasks from your backlog on your own API key.

You get pull requests in your repository, each with its task record, and a weekly report.

  1. 1

    Start: we agree the scope with you, connect to the repository and run a few calibration tasks.

  2. 2

    Every month: you pay for a pool of tasks, and a task counts against the pool when your team accepts it.

  3. 3

    After six weeks we go through the numbers with you: tasks accepted, review rounds, time per task.

You bring: a repository on GitHub, your own Anthropic API key and one person who reviews the pull requests, about 2 hours a week.

This is the proposed shape of a pilot; we agree the details and the price on a call.

Your data and code

The agents work on Claude models (Anthropic), so prompts and code go to the model provider under the terms of the account the work runs on.

Our rule: client work runs on the client's API key or under commercial terms, never on consumer model plans.

Reference project

SmartDrones sp. z o.o. (a related company). We tested the factory at home before offering it. The SmartDrones team uses it to build its own drone-inspection software: firmware, on-board software, cloud and apps. The same team designs the drone's printed circuit boards as code, with AI agents in sessions with an engineer. The code, the designs and the results belong to SmartDrones.

1,203 merges on main since 22 Sep 2026, as of 3 Nov 2026.

A merge is one change, in product code or in the factory itself.

72% of tasks since 15 Oct had two separate reviews (469 of 655 merged on main), as of 3 Nov 2026.

In SmartDrones' board repositories: 2,514 commits co-authored by AI agents since 25 Jul 2026, as of 7 Nov 2026.

About

PromptCompany is a brand of AISolar sp. z o.o., based in Lublin, Poland.

Contact

Tell us what you are building and we will set up a short call about a pilot.