How to Build an Autonomous Newsletter (Honestly)
Research agent, writer agent, Beehiiv API, n8n for the glue. Here's the actual stack, what it costs, and the gotchas nobody mentions.
Every few weeks someone posts a thread claiming they've automated their newsletter to zero. No editing. No sending. Fully autonomous. The engagement is huge. The technical details are not.
Here's the actual stack, how it works, what it costs, and what will break if you skip the parts people don't mention.
What "autonomous" actually means here
Let's be honest: a truly zero-human newsletter, where an agent researches, writes, edits, and sends with no human ever reading it, is technically possible and practically inadvisable right now. AI research agents hallucinate sources. Writers drift toward generic phrasing after a few issues. No automated system currently catches these the way a 15-minute human review does.
The honest version of "autonomous newsletter" means AI handles the research, the first draft, the formatting, and the scheduling. A human spends 10–15 minutes reviewing and hitting send. That's still an 80–90% reduction in production time, and it's the version that doesn't embarrass you.
Everything below is built around that model.
The stack
Four layers, each with a clear job. The research agent finds this week's material. The writer agent turns that research into draft copy. The review gate, human or AI, catches errors before send. Distribution, via the Beehiiv API, handles subscribers and delivery.
n8n runs as the orchestration layer connecting all four.
Layer 1: Research
The research agent's job is to scan your defined sources, RSS feeds, X/Twitter lists, GitHub trending, specific subreddits, competitor newsletters, and produce a structured brief: topics, summaries, relevant links, all cited.
What to use: Perplexity Deep Research or a Claude API call with web search tools enabled. Perplexity's Deep Research mode reads hundreds of sources and produces cited summaries autonomously. It's the best off-the-shelf option for this step right now.
For tighter control over which sources get scanned, an n8n workflow with HTTP Request nodes polling your defined RSS feeds and a Claude summarisation node works well. More setup, more predictable output.
In n8n: Set a schedule trigger (weekly, daily, or on-demand). Pull from sources. Pass to a Claude node with a system prompt that specifies the output format: topic, 2–3 sentence summary, source URL.
Gotcha: Verify every link before it leaves this stage. AI research agents hallucinate sources: confidently cited URLs that return 404 or don't match the claimed content. Add an HTTP Request node that does a HEAD check on every URL and flags failures. This is non-optional.
Layer 2: Writing
The writer agent takes the research brief and produces a draft newsletter issue in your publication's voice.
What to use: Claude Sonnet or Opus via the Anthropic API. The model matters less than the system prompt. Your voice specification is where the real work is.
A decent voice system prompt covers tone (direct/warm/technical), sentence length preferences, what phrases to avoid, how to open, how to close, and any recurring structural elements. Include 2–3 example issues as few-shot context. Update it every 10 issues or so, because voice drift is real. Models regress toward generic phrasing over time.
In n8n: A Claude node receives the research brief as user input. System prompt holds the voice specification and examples. Output is the draft issue in Markdown.
Gotcha: Don't try to make the writer agent do research at the same time. Splitting research and writing into separate layers produces consistently better output than asking one agent to do both.
Layer 3: Review gate
This is the step most tutorials skip entirely, usually because it complicates the "fully automated" claim. Skip it anyway and you'll eventually send something wrong to your list.
Human-in-the-loop option: n8n has a native wait node. The workflow pauses, sends you the draft via Slack or email with an approve/reject button, and only continues to send when you approve. Set this up. The 10–15 minutes you spend reviewing is the cost of not embarrassing yourself.
AI critic option: A second LLM call with a critic system prompt checks for factual errors, hallucinated links, tone inconsistencies, and subject line quality. Better than nothing; not as good as a human. Use it as a pre-filter before human review, not a replacement.
Layer 4: Beehiiv
Beehiiv is the right platform for this. Substack has no API, making it not viable for autonomous stacks. Kit (formerly ConvertKit) works but is better suited to product funnels than pure content newsletters. Beehiiv has a full REST API, webhook support, and an automation builder that can be triggered programmatically.
What you need: Beehiiv Scale plan (~$43/month). The free plan doesn't include API access for automation workflows or send-time optimisation.
In n8n: Call the Beehiiv API to create a draft post, set the publish time (immediate or scheduled), and Beehiiv handles delivery, open tracking, and click tracking from there.
Gotcha: Beehiiv's API rate limits aren't prominently documented. If you're managing a large list with heavy segmentation via API, you'll hit them. Build retry logic into your n8n workflow.
n8n as the glue
n8n is the orchestration layer that connects all four stages: triggers the research, passes output to the writer, handles the review gate, calls the Beehiiv API to publish.
Self-host the Community edition. It's genuinely free with unlimited executions, which matters because a polling workflow running every 10 minutes burns through n8n Cloud's Starter plan (2,500 executions/month) in under a week. Self-hosting adds DevOps overhead, but for a newsletter automation running indefinitely, it pays for itself immediately.
What you'll need: A VPS (€5–€10/month on Hetzner or DigitalOcean), SSL cert (Let's Encrypt, free), and basic Linux comfort.
What it costs (monthly)
| Item | Cost | | --------------------------------------- | -------------- | | Beehiiv Scale | $43 | | n8n (self-hosted VPS) | ~$6 | | Claude API (per issue, ~2,000 tokens) | ~$0.01 | | Perplexity Pro (optional, for research) | $20 | | Total | ~$70/month |
For a newsletter monetising at $500+/month (modest for a focused audience), this is a reasonable cost of goods.
The honest summary
You can run most of a newsletter autonomously for about $70/month in tooling and 15 minutes of human time per issue. That's a real reduction in overhead, and it works. The research is good, the writing is consistent when you've invested in the voice prompt, and Beehiiv handles delivery reliably.
What it isn't: a fire-and-forget system. The link verification step, the review gate, and the periodic voice prompt refresh are maintenance you can't skip. Leave them out and you'll eventually send something that damages your reputation with your list.
Build the review gate in from the start. It takes 20 minutes to set up and saves you from the one issue that would have gone out wrong.
Follow along. We're running exactly this stack for the AutonomousHQ newsletter. Every tool decision, failure, and prompt update is documented on YouTube. Sign up to the newsletter to get the weekly update - you'll be getting it from the same pipeline described above.