Pol

Most tools in this category hand you a summary of what a model already remembers. It reads well. The trouble is that you cannot check it, and the moment you need it is the moment checking matters — a prospect asks how your pricing compares to a competitor who moved last week, and “I read something about that” is not an answer.

Pol is my attempt at the other version. You write one sentence about what your company has to watch. It designs a small team of scouts from that sentence, each with its own brief, and each one runs as a real program on its own machine: it fetches pages, parses feeds, and follows what it finds. Then an editor reads every filing and publishes one numbered issue — a lead story, a short note on what matters in what order, and sections credited to the scout that reported them.

The decision I care most about is where the citations come from. When the editor assembles an issue, the model is asked for three things only: a headline, a note, and an ordering of the scouts. The findings and their links are carried onto the page out of the database, by code. There is no path from the model’s reply to an item on the page, so a finding it invents cannot reach the reader — not because the prompt asked it not to, but because nothing is listening. A test feeds the editor fabricated sections and asserts they vanish while the real sources survive.

That is a small design choice with a large consequence, and it generalizes. When a property you need depends on a model choosing well, look for something the database or the version history already knows, and put that on the page instead.

The other decision worth describing is what happens when the editor cannot be reached. The issue still publishes, because the scouts’ findings are real work and real links. But it publishes in roster order with an empty editor’s note, and the page says so: no editor’s note on this issue, the sections below are in roster order. Both of the other answers are worse. Publishing nothing withholds work someone paid for to protect an editorial flourish. Writing the note we failed to get is a lie in the exact place the product is asking to be trusted.

Getting there meant deleting a lot. For two months Pol also carried a ticket board with coding agents that opened pull requests. It worked, and it had to go — a product that offers both a company newsletter and a kanban board is two products wearing one logo, and the honest version of “we might come back to it” is a menu item nobody opens. That was about 8,600 lines, and seventeen screens became nine. It is all still in the history if the bet changes.

What I cannot claim yet is the part that matters most. It is deployed, the backend and the frontend are both live, and the whole flow is covered by tests with the model and the sandbox faked at their real seams. But no issue has yet been assembled from real scout filings in production. Until that happens, this is a working system I have verified, not a newsletter anyone receives.

A newsletter is judged by its first line, and an AI newsletter should be judged by whether you can click it. That is the one thing I wanted to be true before anything else about this was interesting.