# Software built by coding agents that merges only when the checks pass

AI OS and Baton · Oct – Nov 2026
Author: Henrique Senra (https://henriquesenra.com)
Web version: https://henriquesenra.com/cases/ai-os-vibe-coder-and-baton

> Baton passes each change between a builder agent and a reviewer agent from another model family, and merges only when the checks pass. In one week it recorded 87 review findings in 36 runs, on a product with 1,756 passing tests. I am its only user today.

## Key outcomes

- 87 review findings in 36 runs between 4 and 10 November, 85 from the Mediciners product and 2 from Baton itself.
- 1,756 tests passing, none skipped, with 95% line coverage on the Mediciners CI of 10 November.
- Baton merges a queue run's pull request only when checks and main are green and the product allows it.
- Cost saved, directional: moving rules into code was meant to save tokens. No cost figure exists.

Tags: AI agents, Delivery gates, Evaluation, Hands-on building

## TL;DR

- **Problem.** Coding agents forget what was learned, drift without gates, and spend tokens on decisions a script could make.
- **My role.** I am the product manager, with no technical background to settle technical choices. I decide the product and its perceived quality, and agents execute and check.
- **Decision and cost.** A rule that matters becomes a test or a CI check. Text-only rules failed. The cost is more of my attention spent on the method than on the product.
- **Outcome.** 87 review findings in 36 runs in one week. I am the only user. Business end: cost saved, directional. No cost figure exists.

## Context

When I started working with AI, I realized I had to store what the models learned, or they would forget it. Knowledge kept in one project's folder did not reach another project. From that folder grew AI OS, my set of working rules for every project I run with agents. The rest followed. Vibe Coder is the method I wrote for how I, as a product manager, use AI to build digital products, and Baton is the program that runs it. The method says: "Henrique decides product and perceived quality. Agents execute and check."

For Mediciners, a study product for medical students, I wanted to adapt what I know of product work, a pipeline, kanban, states and specs handed to developers, so it works with agents. It took daily iteration to reach Vibe Coder and Baton.

Baton runs the development stage of the method on a Git checkout. It drives Claude and Codex through their command-line interfaces. One builds and the other reviews, without seeing the builder's reasoning, reading the diff before the plan. Baton merges a queue run's pull request only when the product is set to merge automatically and the checks and main are green. Any other merge is mine.

## What I owned

I do not want to take part in technical decisions, because I lack the background to make them. Agents execute and check. The method lists the only moments that need me: the vision, choosing an item, plan approval, new product decisions, merges outside the Baton queue, the deploy, and checking and closing the release.

## Options

```mermaid
flowchart LR\n  A["Henrique approves the plan"] --> B["Builder agent changes code"] --> C["Reviewer agent from another model family reads the diff"] --> D["CI checks and tests"] --> E{"Checks and main green, product set to auto-merge?"}\n  E -->|"Yes"| F["Baton merges"]\n  E -->|"No"| G["Henrique decides"]
```

| Option | Why not | Cost accepted |
| --- | --- | --- |
| Keep the pipeline in Notion and move cards by hand | The urge to visualize is a wish for control that adds nothing. | The Baton panel shows two things, what needs me and what is running, and not a kanban. |
| Leave the rules as text for agents to follow | Text-only rules have failed. | Each recurring failure has to become a test or a CI check. |
| Run the process through the model | Probabilistic, and spends quota. | Baton in code was created to save token cost, so I write and maintain code. |

Vibe Coder is in Markdown because agents use it. Baton reduces what is probabilistic and increases what is deterministic, for more predictable results and less quota spent. I grow the system by applying many ideas at once and then cutting, and I freeze it for stretches. On 14 November, I decided to deliver a cut and freeze Baton and Vibe Coder again, with this cut criterion: only what stops the Mediciners queue or makes it more expensive, with evidence.

## Results

Baton works well today, in my own words. Between 4 and 10 November it recorded 87 review findings in 36 runs, 85 from Mediciners and 2 from Baton. The Mediciners CI of 10 November had 1,756 tests, all passing, none skipped, with 95% line coverage. Each pull request carries a review and a comparison between the plan and the delivery.

What the numbers do not show.

- I am the only user. The method's own data table lists real product use only for Mediciners, through aggregate counts.
- The release record exists in Baton's code and has never recorded anything.
- Five blocks were installed between 11 and 14 November, and none has been evaluated by use.
- Only 1 of 37 method and Baton items had a build record, so nothing records lessons there.

## Friction and what I would change

The reviewer is supposed to be blind, and it was not. In the review record, the reviewer read the builder's memory in 53 of 147 reviews. The queue also stood for about 33 hours waiting for quota with no record.

I often catch myself changing my thinking and spending queue, tokens and quota on the method instead of on the product. Freezing the method worked well. I plan to buy a Mac to run local models, which I expect to evolve Baton and AI OS and not take them apart.

## Limits

- No cost saving is measured. The token saving is the reason Baton exists in code, not a result.
- The 87 findings are activity, not proof that the software is better.
- One user, and one product accounts for nearly all of the data.
- Clients, colleagues and unreleased products stay private.
