The newsroom: research, checking, writing, art and editing

What I wanted to fix

In the previous post, I covered how I cleaned house, removed duplicate profiles, and kept fewer agents for greater clarity.

Once the setup was in order, the next practical challenge appeared. I wanted to document my progress with technology in this personal journal and teach beginners what I was learning about artificial intelligence. However, sitting down at night to produce complete articles requires gathering facts, writing in two languages, suggesting visuals, and running a strict review so that mistakes never make it into print.

I work during the day, and at night and on Sundays I build with AI agents. My free time is limited. Early on, I tried asking a single general assistant to write a full article about what I had built during the week. The outcome was disappointing. The draft was full of padding, loaded with typical robot clichés, sloppy with factual details, and completely undisciplined. Worse yet, when I asked that same assistant if the text was accurate, it promised that everything was flawless. An assistant auditing its own draft simply cannot spot its own blind spots.

I needed a dependable, repeatable workflow. I wanted each journey post to stay accessible for non-technical readers, have every claim checked against real facts, and go through an editing step that stripped out robotic habits before reaching me. On September 29 and 30, 2026, I decided to rebuild the newsroom from the ground up inside Hermes Cloud.

Quick summary

• The initial problem: asking a single assistant to draft and review its own article creates superficial text full of mistakes it cannot see.
• The solution: on September 29 and 30, 2026, I rebuilt the newsroom in Hermes Cloud, dividing responsibilities among specialized roles: Research, Fact-checking, Writing, Newsroom art, and Tech-editor, supported by R2 and Radar.
• Two core rules: every fact requires two independent sources (or an official primary source), and every draft must pass through a dedicated check and the humanizer filter.
• The style guide: on October 3, 2026, I added the well-explained style guide to keep the writing simple, explain every technical term, and state real costs and mistakes honestly.
• The main lesson: never let the same agent write and inspect its own work. Separating creation from fact-checking is what ensures genuine quality.

First: what an agent newsroom is and what the humanizer does

For anyone seeing this for the first time, talking about an “agent newsroom” might sound like a physical room packed with computers or complicated enterprise software. In reality, it is simply a digital assembly line where each assistant handles one specific chore and hands the work to the next one in line.

Think about a furniture shop. One woodworker cuts boards to exact measurements, another sands the surfaces, a third handles assembly, and a quality inspector checks every joint before packaging. If the same worker who rushed through cutting the timber is also tasked with measuring the finished piece, the odds of overlooking a flawed joint are high, because their eyes are already biased. The same thing happens with artificial intelligence. The workshop analogy helps explain the handoffs, but it stops working at one point: physical parts suffer mechanical wear, while software agents drift when given too many mixed instructions in a single prompt.

The humanizer is a writing skill that works on the assembly line to clean up the prose. Large language models often lean on predictable stock phrases, dramatic hype, and hollow conclusions that instantly reveal computer-generated text. The humanizer scans for these known patterns and rewrites the sentences so that the final reading feels natural, grounded, and conversational, like an honest talk between two people.

What I used

To set up the newsroom in Hermes Cloud, I relied on the following assistants and tools:

• Hermes Cloud: the remote environment that runs around the clock, even when my laptop is turned off.
• Research: an assistant dedicated to gathering facts, dates, and credible sources for each topic.
• Fact-checking: a skeptical assistant that reviews every sentence against actual source data and enforces hard negative rules.
• Writing: the assistant that drafts parallel versions in Brazilian Portuguese and English, keeping the voice simple and human.
• Newsroom art: an assistant that outlines visual concepts and image ideas to match each article.
• Tech-editor: an assistant that structures the final text into the exact layout needed by the site generator.
• R2: the newsroom editor-in-chief, called in whenever an editorial disagreement or tiebreaker is needed.
• Radar: the assistant that monitors the pipeline and delivers the finished draft to my Telegram for review.
• Well-explained style guide: a set of writing principles added on October 3, 2026, designed to explain terms simply, avoid hype, and state real costs and mistakes honestly.

How the pipeline runs day to day

Producing an article for the Journey series follows six clear steps.

1. Fact-gathering with Research. The topic of the day enters the pipeline, and the Research assistant looks only for verified entries in the official briefing. It separates confirmed events from assumptions and organizes the findings into an objective list.
2. Parallel drafting in two languages with Writing. The Writing desk takes the research notes and drafts the article. It creates both the Portuguese text and an equivalent English version sharing the identical structure. The assistant applies the humanizer skill so that the first-person voice sounds like me talking, rather than an automated announcement.
3. Generating visual concepts with Newsroom art. At the same time, the Newsroom art assistant reviews the subject and outlines visual ideas and graphic themes suited to the core message.
4. Structuring files with the Tech-editor. The Tech-editor pulls together the text, metadata, headers, summaries, and sections in the precise format required by the blog compiler.
5. Independent review with Fact-checking. Before any file leaves the newsroom, Fact-checking inspects the draft line by line. It compares every claim against the briefing document. If it finds an unverified date, an invented figure, or a broken rule, it rejects the draft immediately and sends it back for revision.
6. Delivery for human approval with Radar. Once the draft passes the inspection cleanly, Radar sends the full text to my Telegram. Nothing is published to the internet automatically. The post only goes live after I read it, verify it myself, and issue final approval.

What went wrong and what I learned

When I first tested this setup, my biggest mistake was trying to cut corners. I thought I could save effort by putting research, writing, and checking instructions inside a single prompt profile. The result was messy: the assistant claimed it had double-checked everything, yet it happily let fabricated statements slide through. AI models naturally tend to agree with their own reasoning. The moment I assigned Fact-checking its own dedicated profile with the sole job of finding errors and rejecting drafts, quality improved immediately.

Another recurring issue was the stiff, corporate tone of early drafts. Early versions read like company memos, stuffed with grand adjectives and abstract vocabulary that nobody uses in daily life. Pairing the well-explained style guide with the humanizer skill fixed this problem. I had to train the agents to explain technical terms on their first appearance and cut out decorative fluff.

The core lesson from this stage is straightforward: automating text without an independent checking step is the fastest way to distribute mistakes with total confidence. The real benefit of using agents is not the speed of churning out paragraphs, but the discipline of building a system where each step inspects the previous one.

Costs and limits

In financial terms, building this newsroom cost $0 in new expenses. The assistants were configured using the existing Hermes Cloud setup already running for the project.

In time, reorganizing the roles took two evenings of work on September 29 and 30, 2026, followed by adjustments on the morning of October 3, 2026, to integrate the style guide and the daily journey workflow.

The practical limits of this setup:

• Agents cannot replace human judgment. They execute repetitive chores quickly, but deciding what is ethical, accurate, and suitable for the audience rests entirely on me.
• Everything remains strictly a draft. No assistant has permission to publish to the website on its own. The final call is always mine.
• The pipeline takes more time than opening a chat window and asking for a quick reply. That extra processing time is the price you pay for factual accuracy.

What I would do differently

If I were starting over, I would create a separate fact-checking assistant from day one. I wasted hours trying to convince a single writing assistant to catch its own mistakes across long chat logs. Creating assistants with opposing goals, where one creates and the other audits, handles review with far greater rigor and saves time.

Who this is for and who it is not for

Building an agent newsroom with an assembly-line workflow is useful for:

• Anyone who wants to publish educational articles or logs consistently without compromising on factual rigor.
• People studying or building with AI who want to understand how to chain multiple assistants across sequential tasks.
• Anyone looking to produce clean prose that sounds natural and avoids typical automated clichés.

This setup is not for:

• Anyone looking for a one-click shortcut to spit out dozens of low-effort articles every day.
• People who dislike reading drafts and want to outsource final publishing decisions entirely to bots.
• Quick, one-off questions where a simple one-minute chat query gives you what you need.

Wrapping up

Setting up this newsroom in Hermes Cloud gave me the confidence I needed to maintain this public journal, even alongside the demands of my daily job. Knowing that every post goes through dedicated research, natural writing, and strict verification before pinging my phone turns artificial intelligence into a reliable study partner.

In the next post, I will cover Darth: a bot that receives everything and forwards it.

How do you handle reviewing content you build with artificial intelligence? Do you leave it all to one assistant, or do you separate the creator from the reviewer? Let me know on X: @RodrigoesPorto.

Sources and links

• Official Hermes Agent documentation: https://hermes-agent.nousresearch.com/docs
• My official profile on X: https://x.com/RodrigoesPorto