Skip to content

Solutions · Forge

Preview

Agents write the code. A second AI tries to break it.

Forge turns a goal into a written scope, builds each change in isolation and has a second AI attack it before a person approves it. It is how we build VertixOS itself.

See a proposed package before sharing contact details. A person then reviews your answers and sends a scope and figure.

action ledger
  • 14:02:11T1scope.writeexecuted

    goal, files it may touch and checks written first

  • 14:02:40T2build.isolatedexecuted

    its own working copy · nothing else touched

  • 14:19:03T1review.attackrefused

    second AI found a defect after the tests passed

  • 14:31:47T1gates.runexecuted

    full test suite ran · the exit code is the verdict

  • 14:32:10T3promote.mainheld for approval

    waiting on a named person

Illustration of one change. Every row is an event the system records.

What the agents do

From goal to a change you can defend.

  1. 01

    Scopes the job first

    The goal, the files it may touch and the checks that decide the result are written down before any code.

  2. 02

    Builds in isolation

    Each job gets its own working copy, so agents run in parallel without overwriting each other.

  3. 03

    Attacks every change

    A second AI tries to find the case the builder missed, and votes with its evidence attached.

  4. 04

    Runs the checks that decide

    The full test suite and type check run, and the exit code is the verdict.

How it stays safe

Four things it cannot do, however sure it is.

  • It cannot merge its own work.

    A named person approves every promotion to the main branch, and each one is pinned to the commit that passed, so it can be reversed.

  • It cannot review its own work.

    The reviewer is a different agent from the builder. On supported runs it comes from a different model family; that rule is not yet enforced everywhere, and the security pack has the detail.

  • It cannot pass a check that did not run.

    A check that cannot run counts as blocked, never as a pass.

  • It cannot leave its scope.

    An agent that strays outside its listed files has broken its contract, and the record shows it.

See the full list of what agents can and cannot do

Where it stands

Preview

Built and running on our own repository.

214
changes built by agents and promoted
28 June to 18 September 2026
3
model vendors doing the building
29
recorded review rounds
in more than half, the second AI found a real defect after every test had passed
This is our own workload, one operator and not a controlled study. About seven of the 29 rounds used a reviewer from the builder’s own vendor.
Design partnerOn your repository: your engineers bring one real goal and we run it through the whole process, scoped in writing first.

Questions

What people ask first.

It is more than a demo and less than a customer deployment. At 214 changes the process failures surfaced, and one review caught a way for an agent to approve its own high-risk action, before release.

Agent speed is easy. Code you can defend is not.

Five quick questions. A person replies with a scope and figure, or tells you it is not a fit.

See all solutions