Skip to content

IN DEVELOPMENT · NO CHIP HAS BEEN TAPED OUT

AI Employees design the chip

The engineering on BramHA is being set up as the work of AI Employees, each with a role: architecture, logic design, verification, physical design, firmware, and program. They propose. They do not approve.

· 4 min read

IN SHORT

On BramHA, AI Employees are intended to do the engineering. Each holds a role: architecture, logic design, verification, physical design, firmware, or program, which holds the schedule across the others. They write, test, lay out, and re-plan. They do not approve. A named engineer signs the specification, the design freeze, and the tape-out. The program is in development and no chip exists.

A chip program is a set of disciplines that hand work to one another. Someone decides what the chip is. Someone writes its logic. Someone tries to break it. Someone lays it out. Someone writes the software that runs on it. Someone keeps all of that on a schedule.

On BramHA, Element 31's program to build its own silicon for robots, those disciplines are being set up as the roles of AI Employees. The program is in development and no chip exists. What follows describes how the work is intended to run.

The roles

Architecture. This role turns what the chip is for into a structure: which blocks it has, how they connect, and what each must guarantee. For the first design, a real-time controller with a root of trust, that means working out how the control job and the identity job fit in one small part. The architecture role drafts the specification.

Logic design. This role writes the description of the logic that implements the architecture, block by block, in a form the design tools can process.

Verification. This role writes the tests and looks for ways the logic fails to do what the specification says. It works against the specification, not against what the logic designer meant.

Physical design. This role turns verified logic into a layout: where each cell sits, how it is wired, whether signals arrive on time, and whether the result obeys the rules of the manufacturing process.

Firmware. This role writes the software that runs on the chip itself, beginning with the code that starts the part.

Program. This role holds the schedule across the others. When verification finds a fault that sends work back to logic design, the program role re-plans what moves and what waits. The program runs on Tymeline, which holds the date from architecture to tape-out.

Together they write, test, lay out, and re-plan.

Why roles, and not one agent

The division is not decoration. In chip design, the separation between the work that writes the logic and the work that checks it is a safeguard. A designer who tests their own block tends to test what they meant to build. A separate verification role tests what was specified.

Keeping one role per AI Employee preserves that. Each has a defined area, defined inputs, and a defined thing it hands on. The gates between a specification and a tape-out walks through the stages those handovers pass through, and the page at /program shows the roles alongside them.

What they do not do

They do not approve. This is the boundary of the whole arrangement: AI Employees propose, and a named engineer signs the specification, the design freeze, and the tape-out. No AI Employee approves a gate.

It mirrors Element 31's standing boundary for team work: "Agents on a Czar prepare and propose; a named person approves."

On a chip the reason is sharper than usual. Once a design is fabricated it cannot be patched, so the decision to send it is one a person has to own. A named engineer signs every gate explains what each signature covers.

It would therefore be wrong to call the design the product of AI alone. AI Employees do the engineering work. People decide what gets built, when the logic stops changing, and whether the design goes to a foundry.

Where the work is meant to run

The program is being set up so that the AI Employees work on Element 31's own repositories on a Czar Forge, which is Czar configured for technology and AI teams. The intent is that the design stays on hardware Element 31 controls.

That is a plan in progress, not a finished arrangement, and it is written here as intent.

What this does not claim

It does not claim a result. No chip has been taped out. The first design's specification is not published, and there are no figures for how the design performs or how long the work takes.

It does not claim that the method removes risk. A chip designed this way can still be wrong, which is why verification is its own role and why a person signs before the design commits to anything.

And it does not claim much for the first design. It is deliberately small, and a small first design does little. A small design is also a smaller thing for a named engineer to review before signing.

Questions this raises

Do AI Employees make the final decisions on the design?
No. They propose. A named engineer signs the specification, the design freeze, and the tape-out, and no AI Employee approves a gate.
Why split the work into roles?
Chip design is divided into disciplines that check one another, most plainly logic design and verification. Giving each AI Employee one role keeps that separation, so the work that writes the logic is not the work that tests it.
Has this way of working produced a chip?
No. BramHA is in development. No chip has been taped out and no chip exists. The roles describe how the program is being set up.

KEEP READING