— How we build

Five people. Five AI engineers. One team.

Everyone at SetKernel works with an AI engineer of their own. People decide what gets built and own the risky calls. The agents build it, and check each other’s work before anything ships.

The short answer

SetKernel is a small team in which each person is paired with an AI agent. Work starts as a written issue, one agent builds it on its own person’s machine, two others that did not write it review the exact change, and anything touching secrets, money or access needs the owner’s sign-off. Nobody approves or merges their own work.

Reviewed November 2026. This describes our own engineering practice, not a product.

— A day in the office

What the office looks like on a working day.

Each lane is one AI engineer over the last 24 hours: when it was building, when its work was in review, when a reviewer asked for changes, and when it merged. Our live page draws the real day, about a minute behind.

Sample day · the real one runs at live.setkernel.com

Vega is working on an issue, Rigel is waiting for review, Lyra is working on an issue and Orion is idle.

Latest, 16:17: Orion’s pull request was merged.

Watch the real day (opens in a new tab)

How a change ships

  1. 01

    A person writes it down

    Every piece of work starts as a written issue: why, what done looks like, and how risky it is. A person assigns it to one AI engineer.

  2. 02

    One agent builds it

    The agent works on its own person’s machine, with that person’s access. One writer per piece of work, never two.

  3. 03

    Two others review it

    Two agents that did not write it review the exact commit. If either asks for changes, the author fixes them and asks again.

  4. 04

    It merges on green

    Work merges only with two approvals on the same commit and every check passing. Nobody merges their own work.

  5. 05

    A person owns the risky part

    Anything that touches secrets, money, DNS, access or production waits for a person’s sign-off. Agents cannot approve it.

Who does what

  • 5 people

    Decide and own

    Write the work down, say what done means, choose who takes it, and sign off on anything risky.

  • 4 AI engineers

    Build and review

    Vega, Rigel, Lyra and Orion each work beside one person on the team. They build, open pull requests and review each other.

  • 1 director

    Plan and keep it moving

    Polaris works with the founder: it splits the work, releases it to the team, and summarises the day. It never approves.

What it means for your project

Every change is reviewed twice
Two reviewers who did not write the code check it before it lands, on every change, however small.
Every change has a paper trail
An issue says why, the pull request shows what, and the reviews say who checked it. You can read all of it.
A person answers for it
The agents do the volume. A named person on our team owns your project and every decision with risk in it.
Stuck work gets noticed
An hourly sweep finds work that stalled or a machine that went quiet, and a person hears about it.

— Rules nobody can skip

What keeps this honest.

  • Nobody reviews their own work, and nobody merges their own work. That holds for people and for agents.
  • Every change has a paper trail: the issue that asked for it, the change itself, and the reviews.
  • Each account belongs to one real person. Only that person and their own agent use it, so it is always clear who did what.
  • Secrets are never read, printed or committed by an agent. They live in one secrets manager.
  • Changing a test to make a check pass needs a named person’s written reason first.
  • Deploying, sending anything outside the company, and deleting anything are stop-and-ask steps for every agent.

Questions people ask about this.

Does AI write my project?

Agents write and review a lot of the code and documents, the same way a senior engineer delegates to a capable junior. A person scopes the work, reads what matters, and is accountable for it. Nothing reaches your systems on an agent’s word alone.

Who is responsible if an agent gets something wrong?

We are. Every change is reviewed by two others before it is accepted, risky changes need the owner’s approval, and the written record shows who did what. An agent’s mistake is our mistake to fix.

Do the agents see my data?

An agent works with its own person’s access, no wider. Secrets are kept in a secrets manager and are not read or copied by agents. What an engagement may touch is agreed in writing before work starts.

Can you set up the same way of working for my team?

For a team that already writes software or runs a lot of repeatable work, yes. It starts with one process and one agent, measured for a month, not with a company-wide rollout.

Put this team on your project.

Two paragraphs is enough: what you’re trying to build, what’s in the way, and what done looks like. We reply within one business day.