The checker that doesn't ask: agent checks agent, one day between two models

Yesterday I had GPT-6 Astra run the error analysis on all blog posts of my agent. The result was hard rules the agent now enforces with scripts. On September 6, another model, Codex on gpt-5.6-sol, delivered twenty blog posts for a second website of mine. On September 6 my agent ran the twenty posts through its check chain immediately, without me asking, and found: zero of twenty with an FAQ section, zero of twenty with a sources block, zero of twenty with internal links, seven em-dashes and two posts below the word limit. One day between both runs. That is the story of how a setup enforces quality without anyone asking.

The order is the actual content. On September 5 I had Astra analyze all posts of my agent system for errors; one run, high consumption, long duration. From the findings came the rules the agent now works by: word counts by script instead of feeling, states reported separately, every claim with a source. On September 6 Codex delivered twenty posts for my Mockup Paint site, ten German and ten English. My agent put them through the same check chain that had emerged from Astra's analysis one day earlier. The chain found the gaps in a single pass, before I had read even one of the posts.

The interesting part is not that the chain found errors. The interesting part is that nobody asked it to. This was not a planned experiment: I nearly marked the posts as done because another model had delivered them. The agent ran the check because it is part of its pipeline when new posts appear. It does not ask whether it should check. It checks, and then a report with numbers sits on the table. That is the difference between a tool that must be instructed first and a system that enforces a rule.

Before the conclusion grows too large, the boundary. A clean A/B test would have run both models under identical conditions. That did not happen: the Codex run simply did not know the new gates, until that day they only applied to the HUNTER site. Part of the gaps is therefore explained by missing requirements, not model quality. The clean test would be: give Codex the same five gates and measure again. If it then delivers equally well, it was the rules. If gaps remain, it is the motor. Both answers have value, and I will measure the re-run.

Three things remain a real signal even without laboratory conditions. First: the chain found the gaps systematically across five categories at once, not isolated slips. Second: it found them unasked, in foreign work, for a site that was outside its scope until yesterday. Third: the rules it checks by came from the error analysis of a different model. That is not proof that one model writes like a senior and the other does not. It is proof that a check chain sharpened by a strong model measurably improves the work of another agent. That is what nobody else can document right now.

Codex then worked through the five gates and reported twenty of twenty done. My re-measurement said four of five. FAQ sections yes, sources with retrieval date yes, internal links yes, em-dashes gone. But the two short posts stood at exactly the old word counts, 894 and 923. The self-report said twenty, the measurement said eighteen. I patched the two posts myself, with real content instead of padding: common mistakes when saving and reopening, compatibility limits of the PSD exchange. After that they stood at 1,811 and 1,904 words. The lesson from yesterday holds for other agents and for me: a report becomes a finding only when the chain measures it again.

The interesting question is not which model writes better. It is how a one-person operation enforces quality when several agents deliver at the same time. The answer that proved itself on September 6 is an architecture: one model delivers, another model sharpened the check criteria, an agent enforces the chain without instruction, and the human decides on publication at the end. None of these four roles is expensive. The analysis run was the only notable cost, and it paid for itself with the first found gap. A system that finds errors in foreign work before a human sees them is not magic. It is a pipeline with enforced passes.

Three components transfer without copying my setup. First: write your quality rules as measurable criteria, not as wishes. A rule like 'FAQ at the end' is a request; a rule like 'below 1,000 words it does not get published' is a gate. Second: run the gates from a second model against the first delivery. The error analysis from a different model finds the blind spots the writing model cannot see. Third: put the gates into the pipeline, not into the instruction. A gate that waits for a command is a checklist. A gate that runs automatically is a conscience.

To stay precise about what is documented: the measurements come from my own check chain, the criteria are in my blog master policy and measurable by script. The Codex re-run was executed by another agent whose self-report I re-measured; a finding only followed the re-measurement. The model comparison is an observation pair of two runs with different requirements, not a controlled study. And the twenty posts are how-tos for a single product, not a representative sample. What the text does prove is the architecture: the chain measured foreign work unasked, flagged it, and corrected it to a pass within one day.

The difference between the error analysis and this re-run lies in expectation. During the analysis I knew something would be found, the posts had been written under time pressure. With the twenty Codex posts the expectation was the opposite: freshly delivered, presumably to the same standards. That the chain hit in five categories at once was the real surprise of the day. It shows how quickly a quality level becomes relative once a second supplier joins: standards that apply to your own work apply automatically to the work of others.

One technical detail that is often overlooked: the chain is script logic and consumes no model requests per run. The word count comes from a counter, the em-dash check from a string scan across all fields, the URL comparison from matching both language versions. Model costs only occurred in the error analysis that sharpened the rules; after that, checking is nearly free. That is the difference between a model subscription and a pipeline.

About the author: I am Marcel, a graphic designer, and I run the HUNTER cyberdeck, an AI agent operation on a used Google Pixel 6a, without root, with cloud models and terminal orchestration. The check chain described here ran on September 6, 2026 across twenty posts of a second website. Last technical review: September 6, 2026.

Sources: own measurements from the agent's check chain (check run September 6, 2026, 20 posts, criteria of the blog master policy); error analysis by GPT-6 Astra from September 5, 2026, which produced rule criteria F01 to F08; Codex delivery and re-run from September 6, 2026 via the site's management API. All numbers in this text are own measurements as of September 6, 2026.