The Lab

Once a month it takes one repo in rotation and looks for what deserves a deeper design. It brings options; the call stays mine.

A second set of eyes on open PRs: sends them to merge, sends them back with a reason, or writes up a verdict for me. It never touches the code itself — it would be reviewing its own work.

Takes backlog tasks one at a time, test-first, all the way to done, and hands back a draft pull request. It runs until the queue is empty, not until a set hour.

Fires after every “it's done”: re-tests the claim against a fact — a request, a command, a grep — and reopens the task if nothing backs it up.

Reads merged PRs, the dependency audit, and infrastructure drift. Advice only: it doesn't patch and it doesn't merge.

Reads traffic and Search Console, finds where the funnel leaks, and turns each leak into a task. Every finding has to rest on a number.

Every Sunday it turns the week's lessons into deterministic guardrails — and throws out the ones that teach nothing.

Holds the projects up against the written strategy, picks the focus for the week, and opens the tasks that go with it.

Did today's post go out, are the queues filling, did a content pipeline stall. It stays quiet when everything is fine.

Reconstructs the last 24 hours, writes the brief, and sends exactly one ping. It fixes nothing and never touches code.

Walks each product like someone seeing it for the first time and proposes changes. It proposes; it doesn't decide.

Rereads the logs and the task board, then sets next week's priorities. It doesn't code; it thinks.

Fact-checks the draft, then merges or refuses. The last gate before a working professional's public site.

Decides whether it goes out: opens the PR, waits for green, merges.

Waits for CI to go green and merges — or refuses, and says why.

The last gate before publishing: merges up to two articles per run, or leaves the branch alone and says exactly what's wrong with it.

One article and a draft PR. It doesn't merge its own work: another agent checks the facts.

One French article per run, wired into the internal links, delivered on a branch.

One French article per run, delivered on a branch for review.

One article per run, checked against primary sources, in Portuguese and English, internal links included. It never pushes to the main branch.

Checks every week that the promises still hold: buttons go where they say they go, the storefront doesn't contradict the contract, certificates are alive. It only opens a task on a failure it reproduced itself.

A reported error becomes a task for the agents; once it's fixed, it gets closed back at the source.

Watches for a regulation that one product's launch depends on. An agent whose whole job is to wait properly.

Daily digest: disk, memory, containers, updates, backups.

Drafts replies to other people's posts, in my voice. It publishes nothing on its own.

Takes the tasks that need a real browser: visual acceptance after a merge, checks from a residential address. It returns a verdict with its evidence attached.

Checks links a datacenter address can't reach: 403 from the server, 200 from home.

Finds up to five profiles matching one product's target and writes personalized messages. Strict draft mode: it sends nothing and follows no one — I approve the batch.

Prepares and publishes a social account's daily post, once I've approved it with a button.

Checks that the node's browser is alive and that sessions haven't been logged out. It stays quiet when all is well.

The week's numbers on the account, and a read on what actually landed.