Keep evidence at the pace of production
Agents can work in parallel; integration needs a clear record of what was checked and decided.
Industrializing AI-assisted delivery across twelve repositories
This internal project addresses a simple problem: one operator, a dozen products and tools, and agents that produce faster than one person can review.
The aim is controlled delivery: human decisions, verifiable evidence before integration, and visibility over every change.
When agents outpace review capacity
Production speed is useful only when changes remain attributable, verifiable, and safe to integrate without delegating judgement.
The problem to solve
A tooled practice, not a promise of autonomy
The tools organize parallel work while keeping boundaries, validation, and explicit human accountability.
Agent-pass
Parallel agents in isolated worktrees, with validation receipts, provenance for every commit, and a deterministic supervisor.
Tower
A cockpit with no privilege for following work without replacing integration decisions.
Shared instructions
Instructions shared between Codex and Claude, checked by a CI freshness check.
Curator passes
Passes that check the backlog against the code before implementation.
Observable results, without a percentage promise
The figures describe the internal practice; they are not a projection of gains for a client team.
Test growth
Emovanti's test code grew from about 12,000 to about 202,000 lines in seven months.
The four published markers
They describe the scope and discipline of this internal practice.
- Commits
- About 5,450 in eight months across twelve repositories.
- Branches
- 458 agent branches in Emovanti alone.
- Tests
- About 12k to about 202k lines in seven months.
- Decision
- One operator who decides and reviews.
What this study establishes
An agent-assisted delivery practice must stay explainable and controlled at every integration.
A single-operator practice, not yet a team rollout.
This experience proves a method on LRJI's internal projects. It does not claim to describe a client rollout or an already deployed team.