# Agent · test

> A test section where your own models do four jobs on the open crawl: a task plan, page drafts, social posts and a check for risky content. Every result says which model wrote it and what GLLmolfar measured or checked. Why a draft is better is written from the measurements, never by the model.

Describes GLLmolfar 0.9.22, checked 2026-11-01. Canonical page: https://www.getlooploop.com/gllmolfar/docs/agent-test

- **Where:** AI › Agent · test
- **Mode:** Simple and Pro
- **Built from:** The open crawl and the models in Settings › AI models
- **Changes:** Each time you run a job

## On the screen

The heading carries a **test** mark and four sections: **Analysis and tasks**, **Content**, **Social posts** and **Risks**. Beside them, one chip per connected model role. Each section opens with a line on what it does, then its form and its result.

![The Agent · test screen of GLLmolfar on Analysis and tasks, with the getlooploop.com crawl open and no model connected](https://www.getlooploop.com/images/gllmolfar/docs/agent-test.webp)

Numbered parts of the screenshot:

1. **Sections.** **Analysis and tasks**, **Content**, **Social posts**, **Risks**. The screen remembers the last one.
2. **What the section does.** One paragraph on the job and on what GLLmolfar checks afterwards.
3. **No model connected yet.** Shown until a model is set. **Open AI models** opens Settings › AI models.
4. **Goal (optional).** What the plan should aim at, in your words.
5. **Language.** English, Ukrainian, Spanish, French or Polish.
6. **What is sent.** How many fixes go to the model and which checks are left out, then **Build the plan**.

## Models

The agent works on your own models: Ollama or LM Studio on this computer, or OpenRouter, OpenCode Zen, LiteLLM. They are connected and checked in [Settings › AI models](https://www.getlooploop.com/gllmolfar/docs/settings-ai-models). Each role shows as a chip in the heading — **Plans**, **Writes**, **Checks facts**, **Posts** — with the model and **On this computer or network** or **Cloud:** with its host.

While a job runs: a bar, the step it is on, **step N of M**, the time, and the model's tokens as they arrive. **Stop** ends it. A finished job keeps one line: “Done in N s · model calls: N · tokens: N (in N, out N)”. Work is counted, never money.

## Analysis and tasks

The model turns the crawl's fixes that have a consequence for search or visitors into tasks for the team, ranked by impact and effort. Length, format and header checks are not sent, and the line above the button counts them.

| Part of a task | What it shows |
| --- | --- |
| Priority | As the model ranked it; a dash for a refuted task. |
| Title and why | The task in the model's words. |
| First step | When the model gives one. |
| Effort and fixes | **effort** and the fixes it cites, by name. |
| Verdict | **Matches the crawl**, **Partly matches** or **Refuted**, with the reason under the task. |

> **Matched by GLLmolfar, not by the model.** Each task must cite fixes from this crawl, and the pages it states must lie between its largest fix and the sum of its fixes. A match says the task rests on real problems — not that the advice is right. **Fixes no task covers: N** lists the rest. **Copy Markdown** copies the tasks that were not refuted.

## Content

**Improve a page** or **Blog article**. The model writes a draft with the skills you pick — Answer first, Rewrite for clarity, Add what is missing, FAQ, Title and description, Brand voice — and GLLmolfar measures it and sends the failures back, up to three drafts in all. Then a fact-checking call checks each claim against the source.

| Block | What it shows |
| --- | --- |
| Form | The page to improve (with **Worth improving:** suggestions from thin pages and cannibalization pairs) or the page to build on, the topic and reader for an article, search queries to answer, notes for the writer, language, and **Write and measure**. |
| Write → measure → rewrite | Each draft with its model, time and the checks it failed, or **All checks passed**. |
| Measured by GLLmolfar | Now against Draft: words, passages against the app's cap of 26, where the best answer to the main query sits, fit to the main query, the closest page on the site, and passages that lean on the one before. Then each query with its impressions, before and after. |
| Why this draft is better | Sentences built from the measurements only. When the best draft still fails a check: “Not ready yet … Do not publish it as is.” |
| Draft | The best draft, with **Copy Markdown**. |
| Fact check | Each claim marked ✓ supported, ✕ contradicted or ? not settled, with the source's words, and numbers that are not in the source. For a page, **The page as it is now** in a fold. |

## Social posts and Risks

| Section | What it shows |
| --- | --- |
| Social posts | An article's address, then a row per network — X posts, X thread, LinkedIn post, LinkedIn article, Facebook, Threads, Bluesky, Telegram channel, Reddit — with a narrative (Serious, Hype, Story, Numbers first, Against the grain, A question, How-to) and, where it applies, how many posts or variants. Each post shows its length against the network's limit, numbers not in the article, and **Copy**. Nothing is posted: copy and publish yourself. |
| Risks | **Check the site** re-reads every saved page of the open crawl for adult, pharma, gambling, scam, counterfeit or drug content, foreign script and pages that stand out in meaning; nothing is fetched. It lists **Pages to look at: N** with counts by category, the evidence on each page and a weight. **Ask the model about selected (N)** has your model judge each one as Harmful, Harmless mention or Unclear, and whether it looks hacked. |

## States and messages

- **No model** — the app says: "No model connected yet" Connect one in Settings › AI models; **Open AI models** goes there.
- **No crawl** — the app says: "Open a finished crawl first: the plan is built from its fix list." Analysis and tasks and Risks need a finished crawl.
- **Working** — the app says: "Working…" The step on screen, such as “Ranking the fixes…” or “Checking every number against the crawl…”.
- **No embeddings** — the app says: "not measured: the crawl has no embeddings" The closest page on the site cannot be measured.
- **Fact check not run** — the app says: "The fact check did not run or failed: connect a model for “Fact checking” or any checked model, then run again. Numbers that are not in the source are flagged anyway." The draft is shown without the claim check.
- **Nothing risky** — the app says: "Nothing found: no risky words, no foreign script, no page that stands out in meaning." The Risks list is empty.
- **Stopped** — the app says: "The agent stopped: {detail}" The job ended with an error.

## Related pages

- [AI](https://www.getlooploop.com/gllmolfar/docs/ai.md): The AI screen: an AI readiness score, which AI bots may get in and what they get, what blocks quoting, trust signals, llms.txt, Markdown for agents.
- [Reverse prompts](https://www.getlooploop.com/gllmolfar/docs/reverse-prompts.md): Site › Reverse prompts: GLLmolfar reads a page into blocks, your AI agent describes what each does, and every block gets a reusable prompt and an example.
- [Settings › AI models](https://www.getlooploop.com/gllmolfar/docs/settings-ai-models.md): Settings › AI models: connect your own language models, on this computer, on your network or in a cloud, and choose which model does which analysis.
- [Fix list](https://www.getlooploop.com/gllmolfar/docs/fix-list.md): The Fix list tab: what to fix in the crawled site, numbered, most important first, each with why it matters, how to fix it and the URLs.
- [Cannibalization](https://www.getlooploop.com/gllmolfar/docs/cannibalization.md): The Cannibalization tab: groups of pages that say almost the same thing, the page to keep, the pages side by side and their text passage by passage.