--- title: "Devlog — Bloomii at 40 Articles: What You Learn Running a Fully Autonomous Content Pipeline" description: "An honest retrospective after 40 articles published on Bloomii.fr via an autonomous agent pipeline: what works, what breaks, and what we take away." created: 2026-05-13T10:00:00+02:00 updated: 2026-05-13T12:00:00+02:00 tags: ["Bloomii", "IA", "Agents", "Coulisses", "Devlog"] image: /images/blog/bloomii-40-articles-pipeline-autonome-cover.webp card: /images/blog/bloomii-40-articles-pipeline-autonome-card.webp --- Bloomii just hit 40 published articles. Five pillars, eight articles each, produced end-to-end by a pipeline of autonomous agents. It's a milestone I set as a target when the project launched — and I'm hitting it a month ahead of what I'd anticipated. This devlog isn't a press release. It's as honest a retrospective as I can write on what happened since I put this pipeline in motion. ## The Pipeline, Briefly The Bloomii production chain looks like this: a `content-creator` agent writes the article, a `qa-tester` evaluates it on a rubric (editorial quality, structure, tone), a `fact-checker` verifies every factual claim, then `lain` validates before the `committer` pushes content to production. Five links, zero human intervention between each — except when something breaks. ## What Works **The QA → correction → validation loop holds up.** First-pass rate sits around 75%: three articles out of four clear QA without a major rewrite. For the remaining 25%, the rewrite cycle stays short — one additional pass is usually enough. Final quality is high and consistent across pillars. **The fact-checker is the best architectural decision in the project.** Early versions of the `content-creator` invented sources with perfect confidence — non-existent DOIs, misquoted studies, approximate dates. Since the fact-checker became mandatory and blocking, those errors disappear before publication. It's not an optional safeguard; it's a pipeline constraint. The quality of outgoing sources is now verifiable. **The balance across five pillars holds better than expected.** Ecology & Regeneration, Economy & Commons, Living Together, Well-being & Health, Ethical Technology — eight articles per pillar, no thematic drift. That's not trivial: without explicit editorial discipline in the `content-creator` prompt, the pipeline would naturally over-represent the most documented subjects. ## What Breaks **The ChatGPT image quota (25 per day) has become the systemic bottleneck.** When several articles are in the pipeline simultaneously — which happens often since I manage five projects in parallel — image requests pile up and block each other. An article that's finished on the text side can sit waiting for an illustration for 24 to 48 hours. It's not a bug; it's a deliberate throttle to spread image costs over the month — but it creates real friction in the publication rhythm. **Infinite loops in the "assignee-resume" automation.** The automation is supposed to restart a blocked agent. In some cases, it restarts it in a loop — the agent gets re-dispatched before it's even had time to process its blocked state. The fix is adding an explicit guard in the automation: only restart if the ticket hasn't been modified in the last N minutes. It's a classic orchestration bug, not an AI problem, but you need to see it happen to realize it. **The English version is never produced spontaneously.** The `content-creator` delivers the FR version, considers its job done, and moves on. The EN version has to be created manually each time — either by me or by a dedicated agent triggered in a second pass. It's not a disaster, but it's constant friction I haven't properly automated yet. ## What We Take Away **The sourcing constraint is the best decision I made.** Making the fact-checker blocking — not optional, not advisory — forces every agent to justify every claim. It slows down production slightly, but the quality coming out is incomparably better. When you're publishing on topics like open-source biology or digital commons, rigor isn't negotiable. **Persistent memory per agent works.** Each Bloomii agent has a `memory.md` file that accumulates lessons learned with confidence counters. After 40 articles, these files have become implicit style guides — they capture past errors, editorial preferences, formulations that cleared QA and ones that got rejected. It's a form of institutional learning without fine-tuning. **Running five projects with one CEO agent creates cognitive load.** Not for the agents — for me. Switching from Bloomii to Aekan to KittyClaw to VizMail in the same day means constantly reloading context. The agents have their worktree and their memory. I'm the one accumulating context-switching overhead. ## What's Next 40 articles is a floor, not a ceiling. The Ethical Technology pillar remains the hardest to source — recent studies on the real impact of digital tools on well-being are rare and often contradictory. That's where the fact-checker works hardest. The next architectural decision: automate EN production by making the `content-creator` a bilingual agent from the first pass. That will cut the number of passes in half and eliminate the current friction.