AI Agent Auto-Publishing: Brakes Beat Gas
A blog pipeline that researches, writes, verifies, and publishes on its own, plus the four gates that make it trustworthy.
TL;DR
The author's blog publishes articles fully automatically via a 2-hour cron pipeline that researches, drafts, verifies, and posts without human touch. Safety comes from strong brakes: verification gates that reject bad content, automatic rollback after publishing, and escalation to humans for taste calls. Full-auto is trusted only for mechanical tasks; editorial judgment stays human.
At 6 AM a notification arrived. The cron job on adityo.web.id had just finished its cycle. An article was already live, neatly: deep research, perfectly formatted code blocks, valid embed tokens, and a synchronized English translation. Without me touching anything.
There was a real mix of pride and nervousness in that moment. But this is not magic. It is the result of a pipeline whose "brakes" were built far stronger and more carefully than its "gas".
Pipeline anatomy: a workflow, not a rogue agent
This blog runs a cron every 2 hours that operates fully without human intervention. The flow is strict and measurable:
# 2-hour cycle on adityo.web.id
monitor_commit_trend # 1. Watch latest commits/trends
research_ledger # 2. Research: min. 3 official sources opened, verbatim quotes recorded
decision_gate # 3. Decide: worth writing or not?
llm_draft # 4. Generate the article draft via an LLM
verify_gate # 5. Auto-verify (SEO, code blocks, tokens, AI artifacts, EN sync)
publish # 6. Publish to live
watchdog_reverify # 7. Post-publish reverify; fail = auto rollback to draftIt is important to be honest: this is not a "fully autonomous agent" roaming free. Referring to Anthropic's research, "Workflows are systems where LLMs and tools are orchestrated through predefined code paths". This pipeline is honestly a workflow, not a rogue agent, because workflows give "predictability and consistency for well-defined tasks". We need determinism here, not surprises.
Four pillars of safety
One, trust is granted gradually. In the beginning, every publish required human approval. After the flow stabilized and the verification gates were tested, manual approval was removed in favor of full auto. But a compensating control is mandatory: a watchdog that reverifies right after publishing, ready to trigger an automatic rollback on any anomaly.
Two, verification must be able to reject. The verification gate keeps a list of FAIL (blocking) and WARN (non-blocking) conditions. FAIL covers broken structure, zonked embed tokens, translations that shift code block contents, or chatbot conversation artifacts. If a FAIL is detected after publishing, the rollback from published to draft happens automatically. That is not a hope; it is a hard system rule.
Three, distinguish mechanical from non-mechanical fixes. If verification fails over something trivial (a typo, an empty SEO field, messy whitespace), the evolver agent may fix it itself and push again. But if the problem is the writing angle, questionable facts, or narrative structure, the system stops and sends an ALERT to a human. Agents have no taste, and must not be given the authority to judge taste.
Four, content safety built on real sources. Research must open at least 3 official sources. Raw numbers from search snippets are forbidden in drafts. Every key number or claim must be hyperlinked to its original source. This aligns with Google's guidance emphasizing "first-hand expertise and a depth of knowledge (for example, expertise that comes from having actually used a product or service)". Automated content without quality control is high risk, because Google detects violations "both through automated systems and, as needed, human review that can result in a manual action" per its spam policies. Automation plus quality gates plus real experience is the only safe path.
Full-auto does not mean brakeless
My opinion is firm: building a full-auto system is not about making it run as fast as possible, it is about making it stop safely when the unexpected happens. The "brakes" in the form of verification gates and automatic rollback are the most important part, and ironically they take the longest to build.
So when do I trust this agent? When its tasks are mechanical, structured, and repetitive. What I still never delegate: judging the feel of the writing, validating untested new facts, and final editorial decisions. Let the machine be a reliable machine, while humans stay thinking humans.