Issue 25 went out on Monday. This special is not a recap of the stories. It is about the system behind them: how it is built, which integrations it needed, where it failed, and what I still decide myself.
A quick reminder first: Bits & Orbits is a personal R&D project, separate from my work at ESA. Nothing here comes from ESA-internal information.

Why I built it
I'm a space engineer, and I follow every new high-tech development in the field. Late last year I also started looking more closely at generative AI and LLMs.
It became clear that the two are converging: orbital data centres built for AI on one side, AI-driven spacecraft autonomy on the other.
I decided to investigate that intersection privately. Rather than write the research by hand, I wanted to use AI itself and build an agentic environment around it. That became Bits & Orbits, and it turned out to be one of my better decisions. It taught me as much about using AI as an operating system as about the space-AI frontier I set out to understand.
The first issue
My picture at the start was simple: information goes in, a model writes, a checker checks, the platform publishes, and I approve at the end.
That picture broke before the first issue. The pre-release issue only reached you because I pasted it into Beehiiv by hand. The automated publish step said "success" and did nothing. It had fallen back to an API call that was refused, and the workflow still showed green.
When I read the documentation properly, I found that the publishing path I had designed around isn't available at my tier.
So I removed the retries, added one manual paste step, and made the workflow fail loudly if that step is skipped. That is still how every issue reaches you. The LinkedIn cross-post is manual on purpose too.
First lesson: an honest manual step is better than an automated step that reports success when nothing happened.
What runs today
The pipeline today has these stages:
A source monitor reads the feeds I've chosen and remembers what it has already seen.
A scout searches the web for what those feeds miss, once for AI for space and once for space for AI.
A prefilter removes stale and off-topic items before any model spends time on them.
A triage step decides what is worth a closer look and records why.
A dedup step makes sure we're not repeating an earlier issue.
A drafter writes the draft.
A fact-check gate traces checked claims back to their sources.
A scheduler sets the send time.
A sprint journal records what changed in the system that week. That becomes the Meta-beat note linked in each issue.
None of it was designed on day one. Almost every stage exists because something went wrong without it.
Most of the work was in the integrations. Feeds, web search, the language models, version control and CI, Beehiiv, LinkedIn, the public Insights dashboards and the Meta-beat page all have to talk to each other. Each connection failed at least once:
A web search lost its place when it paused and resumed.
A clustering step broke because it wasn't streaming a long model response like the others.
A configuration error dropped two candidates before anyone had read the full article.
Most of the "AI work" turned out to be plumbing.
Failures that looked like success
Crashes were never the real problem. A crash tells you something is wrong. The harder failures were the ones that reported success.
In a single week, the search step returned nothing and raised no error, a story I had added by hand became invisible because of a mangled link, and another step overwrote my reviewed selection with the raw one. Three bugs, all reporting success.
Another week, the journal that summarises my work didn't have permission to see it. It looked at one of my busiest weeks and concluded that almost nothing had happened. I nearly published that.
The scheduler
In Issue 21 the scheduler queued the issue a week late, because a Monday date in the title was treated as already past the send day. I added a pin and a test.
The next week the bug was back, and the test hadn't caught it.
Then Europe switched to winter time. A send time fixed in UTC would have reached you an hour earlier than usual. We caught that shortly before scheduling Issue 23 and changed the scheduler so a clock change doesn't move the delivery hour.
This Monday, Issue 25 went out on the automatic local-time calculation, with no manual adjustment. Making "Monday morning" boring took longer than any of the AI parts.
Prompts and rules
For three weeks I wrote ever more emphatic instructions telling the drafting model to cite every paragraph. About a quarter still came back uncited.
What finally worked wasn't a better prompt. It was moving the rule out of the instructions and into plain code that stops the run. A prompt is a request. If something must always be true, it needs a gate.
Where the value comes from
I spent real time choosing good sources. Then, at one point, they contributed nothing to the issue for five weeks in a row, while open search carried it.
Search is still doing its job. This week a scout found the Ubotica CogniSAT-6 story you read on Monday.
Some feeds mostly add noise. In Issue 24 one press-release index supplied 26 of 29 sampled drops, all without a publish date, and I don't yet dedupe across sprints. This week 13 of 20 sampled drops came from just two feeds: one mostly event promotions, the other again with no AI angle.
And one more: once I re-ran the planning stage mid-week to pull in two late stories. Because the system only saw what was new, six stories became two. Since then, a refresh merges with what is already there. It never regenerates.
Where the human stays
Every week the fact-check comes back green, and I still find things in review. A navigation story with no machine learning in it. A hardware "first" I couldn't stand behind. A story whose AI angle was entirely "will support" and "soon".
This week I removed two repeats in brief review: Google's Project Suncatcher, which we had already covered in Issues 19, 20 and 24, and a NASA navigation flight from Issue 19. The sources were accurate every time. It was on me to make the judgment.
I see my role as three jobs. I'm the editor, deciding what is worth saying. I'm the operator, keeping the machinery honest. And I'm the sign-off, the person who answers for what goes out. So far, every word that reaches you is still reviewed by me alone.
If you read Monday's lead story, you saw the same idea from the space side. Orca Aerospace described how its onboard autonomy software for Orbital's planned data-center satellites will work: shadow mode first, with telemetry sent down for human review, then autonomy in layers. A small AI model suggests, a rules-based software turns that into commands, and a safety kernel checks each command against the limits the operator sets.
Now swap “satellites” for “newsletter issues” and that is roughly where I've landed. The model proposes, rules enforce, a human signs, and autonomy is earned one layer at a time.
I don't think any of this is really about newsletters. Swap "issue" for a research summary, a weekly ops report, a product decision or a customer reply, and the same patterns show up.
I now see an AI-native organisation less as one where AI does the work, and more as one where the work is described precisely enough that machines can do parts of it. Every step leaves a trail. Rules that matter become checks instead of habits, and those checks cover what people actually receive. And people sit deliberately where judgment and accountability are, not wherever there happens to be a gap.
That takes more structure, not less. Writing the Meta-beat in public every week is what forced that structure on me.
7 things I would do again
Make failure loud. "No errors" is a claim to verify, not a result.
Put hard rules in code. A prompt is a request. If something must always hold, enforce it outside the model.
Trust a test only after it has caught the real bug.
Flag silent loss. Output that shrinks, drops a section or comes back empty should raise a flag.
Make re-runs and handoffs safe. Repeating a step merges. It never silently starts over.
Keep a human at sign-off. People sit deliberately where judgment and accountability are
Log what the system did, every cycle. In public, if you can.
And finally, running Bits & Orbits usually takes me under two hours a week, and about €40 a month in tokens.
Thanks for reading, whether you've been here for all 25 weeks or just this one. If you run an AI pipeline yourself, at work or on the side: which step do you still check by hand?
All issues: https://read.bitsandorbits.com/
The behind-the-pipeline log: https://bitsandorbits.com/meta-beat/
Subscribe: https://bitsandorbits.com/subscribe/
