AI agents that write and check fight analyses for a subscription business

We built and run Crenas Coliseum, a subscription that publishes an analysis of every UFC fight before each event, 12 to 14 a week, in Portuguese and English. Here is how the agents write and check them, and what we learned.

The starting point.

12 to 14 analyses every week

Each week brings a new batch of analyses, all due before the same event.

Two languages

Every analysis is published in Portuguese and English, written for each one, not translated line by line.

Every fact has to be right

Subscribers pay for the analyses, so a wrong number costs trust.

A small team

The founders run the business, so the work had to be done by agents and checked by rules.

01

A team of agents, each with one job.

Four of the business's seven agents do the weekly writing: one finds what is due, one writes each analysis in Portuguese and English, one checks the facts and one fixes what failed. A simple program, not the AI, publishes the result.

Step 1Find what is dueWaitingAI agent
Step 2Write, in two languagesWaitingAI agent
Step 3Check every factWaitingAI agent
Step 4Fix what failedWaitingAI agent
Step 5PublishWaitingProgram

Example figures. The steps and the rules are the real ones.

The AI never publishesThe AI writes, but only a simple program can publish, after checking nothing is missing.
A summary email every weekThe team gets an email with what went out, what was held and why.
02

Every fact is checked against two sources.

A separate checking agent can read but not change anything. It confirms each fact in two places, and anything it cannot confirm is held, never counted as passed.

Analysis draft · Fight 7Example
  1. Fighter's record18–4Source ASource B
  2. Last fight12 SepSource ASource B
  3. Wins by knockout9Source ASource B
  4. Ranking in the division4thSource ASource B
Rule for the checkerWhen in doubt, it fails.

A fact counts only when both sources agree.

Rewrites used on this analysis0 of 3

Each rewrite goes back to the same checker.

Read-only checkerThe checking agent can look things up but cannot edit the analysis or anything else.
Held, not guessedA fact that cannot be confirmed holds the whole analysis back until it is fixed.
The same checker every timeA rewritten analysis is checked again from the start, not waved through.
03

Nothing goes out unless at least 85% of the week passes.

If fewer than 85% of the analyses pass, or the main-event analysis fails, the whole batch goes back to draft and the team gets an email. The rule came from a real week when too few analyses were ready.

Example figures. Analysis 1, with the dashed border, is the main-event analysis.

Goes out85.7% passed12 analyses go out. The 2 that failed stay in draft.
The whole batch, or nothingA half-finished week never goes out: below 85%, every analysis waits.
The main-event analysis decidesIf the main-event analysis fails, the batch is held even when the others pass.
An email explains itThe team sees what was held and why, and decides what to do next.
04

At most 3 rewrites, and no changes in the last 3 days.

An analysis rewritten too many times can drift from what was asked, so each one gets three tries. Three days before the event, everything is frozen, and after it a review adds lessons to the writing rules.

Example analysis. Blue days are frozen: nothing changes.

Analysis · Fight 3Open for changes
  • Draft written and checked
Three tries per analysisAfter the third rewrite, an analysis that still fails is held back instead of changed again.
Frozen three days beforeSubscribers read a stable analysis in the final days, with no late surprises.
A review after each eventEvery analysis is checked again afterwards, and the lessons go into the rules the writers read.

The project in numbers.

12–14fight analyses a week, written and fact-checked by AI
85%of a week's analyses must pass before anything goes out
2sources for every fact
3rewrites at most per analysis, and 3 frozen days before the event

Source: the code and settings of the system that writes the analyses, 21 October 2026.

What we learned.

01The AI writes, a simple program publishes.

Keeping the agents away from publishing makes every change easy to check and easy to undo.

02Hold what can't be confirmed.

A fact with one source waits for a second one, so subscribers only see what was confirmed.

03Set a rule for the whole batch.

The 85% rule keeps a weak week from going out half done.

04Limit the rewrites.

Three tries and a freeze before the event keep each analysis close to what was asked.

05Keep people in the loop.

The founders read what goes out, and the weekly email tells them exactly what was held and why.

← Back to AI agents Want agents for your team's routine? Tell us about the job →