How it works

Five phases.
Zero shortcuts.

Each phase has a dedicated agent and a gate. Nothing advances until the work meets the bar. You step in when you want to. Everything else runs itself.

PHASE 01PlanPlanner
PHASE 02DesignOO Design Gate
PHASE 03ExecuteExecutor
PHASE 04ReviewBughunter · Architect
PHASE 05ShipPR Reviewer

Plan — before anything moves.

Work starts with a conversation, not a prompt. Veriloom's planner guides you through breaking your project into features, epics, and tickets. Each gets a clear spec and acceptance criteria. Nothing moves forward until the scope is real.

01
Conversational scoping, not one-shot prompting.
02
Tickets come out with spec and acceptance criteria attached.
03
Vague work is sent back. No ticket advances without a real definition of done.

Design — the OO gate.

Before a single line of code is written, each ticket passes the OO Design Gate. Object-oriented principles, proper abstractions, and clear boundaries are verified. Vague tickets go back. Approved tickets go forward.

01
Abstractions reviewed before code is written.
02
Boundaries, dependencies, and responsibilities explicitly checked.
03
If the design is unclear, the ticket returns to Plan.

Execute — one agent, one worktree, test-first.

A dedicated agent is launched per ticket, inside its own isolated worktree, with full project context. It works test-first. When its context window approaches the limit, it writes a handoff summary and a fresh session picks up exactly where it left off. Nothing is lost between sessions.

01
One agent per ticket, in its own isolated worktree.
02
Test-first by default. Tests come before the implementation.
03
Context-window handoffs preserve state across sessions.

Review — two independent checks.

Finished work runs through two independent checks: a bug detection pass followed by a separate architectural review. Both must pass. If either fails, the report is sent back to the executor automatically and the loop runs again.

01
Bughunter does a structural pass. Then, separately, the Architect reviews.
02
Two gates. Two independent opinions. Both must clear.
03
Failures loop back to Execute with the specific findings attached.

Ship — you merge when you're ready.

Once a ticket clears both reviews, a pull request is created and queued. A dedicated review agent monitors the main branch every ten minutes. You merge when you are ready.

01
PR opened automatically with full context and test results.
02
Main branch monitored every ten minutes by a dedicated reviewer.
03
Final merge is always yours. Nothing lands without a human click.

Ready to run real engineering discipline?

Join early access and see the process in your own codebase.

Get Early Access →