Six AI agents, one release, checked twice

Parallel work only pays when the pieces do not collide, so the first job was cutting the list into scopes with separate files:

• a full capture test against a small practice website;
• brand cards that stay readable on any site colour;
• the call to action speaking the film's language, and a grader that stops marking down the cursor;
• site plans kept per language, with active sites scanned again every week;
• clearer status words for the MCP tools, and a test that weighs the tool list;
• a nightly run of the end-to-end suites.

Each agent worked in its own copy of the repository, with a brief naming its files and one instruction above all: verify your work before you report. All six finished in ten to nineteen minutes each.

The most valuable paragraphs were the ones where an agent told us our brief was wrong. One found that the component we blamed for a contrast problem never called the function we had named. One showed that our size target for the MCP tool list was smaller than its bare structure. One showed that only one of our three end-to-end suites could run against the live product at all.

That is the real gift of giving a careful agent a narrow job. It reads the code with fresh eyes, and it has no stake in the assumptions you wrote into the brief.

The capture test serves a small website from the machine and films it with the real recorder. A sub-page that answers with an error once is filmed again; a sub-page that sends you to a sign-in falls back to the main page. And when the main page itself stays unreachable, the film is now declined clearly and its credit comes back, instead of going ahead without footage.

The brand-card change is just as concrete. A light site often owns a light brand colour, and white type on lime measured only 1.7 to 1 in contrast. Brand cards now carry a steady shade behind the type that brings white text to at least 4.5 to 1, even over pure white, while films of dark sites do not change by a single pixel.

We merged the six branches one at a time, then ran the full gate once on the combined code: type checks, lint, the complete test chain, 101 suites and 4,681 checks that night, and a production build.

The second verification was ours, not the agents'. Each agent had only ever seen its own branch, so we ran every agent's own tests again on the merged code. Two small frictions are now written into the brief from the start: a fresh copy of the repository has no installed dependencies, and every agent that adds a test edits the same line that chains the tests together, so that line is merged by hand.

One scope deserves its own line. Our three end-to-end suites came at the end of the test chain and stepped aside whenever their environment was not set. Now they run every night against the live product.

The MCP suite makes only calls that are declined before any credit or record is created, so it costs nothing and changes nothing. If it is ever missing what it needs, the run turns red and names exactly what to add. A skipped check is now one visible line, never a silent pass.

You can read more about the MCP side of this work on our assistants page.