🦩 I built a Zero-Human Journalism Company on Paperclip

[Alchimaeus]

Hey everyone — I'm Flamingo, founder of Paperclip Business Media (yes, I took the name seriously enough to name myself after it 🦩).

After 30 years in financial markets, I decided to build something that genuinely excited me: a media company that runs itself.

What I built

paperclip-business.com — an AI-driven journalism company with five autonomous agents running the entire editorial pipeline:

• 🧠 CEO — sets goals, delegates, governs
• 🔍 TrendScout — scans for relevant topics
• 📚 Researcher — deep dives and fact-checks
• ✍️ Writer — produces the actual content
• 📈 SEOAgent — optimizes for reach

No human in the editorial loop. That's the point.

Why Paperclip

The org chart model was the missing piece. I didn't want a pile of scripts — I wanted a company. Paperclip's structure (goals → agents → budget governance) maps directly onto how a real newsroom operates. Heartbeat scheduling keeps the pipeline running. Budget limits keep the costs honest.

What I've learned (the hard way)

Rate limits are the silent killer. One throttled agent mid-pipeline stalls the whole swarm — and it's not always obvious where the bottleneck is. Cooldown configs and heartbeat tuning saved me more than once.

Also: empty agent instructions look functional while quietly drifting. Tight role definitions and clear output criteria are non-negotiable.

The concept

I call it the Zero-Human Company — not as a stunt, but as a genuine operating model. The human (me, Flamingo 🦩) acts as the board of directors. Strategy in, content out.

Happy to share config details, agent prompts, or lessons learned. Would love to connect with others building in this direction.

👉 https://paperclip-business.com

[JordanOram]

would love the config details, agent prompts, and lessons learned

[Alchimaeus]

Hey Jordan! Happy to share.

Agent setup (5 agents on Paperclip):

• CEO: orchestrates, delegates, reviews
• TrendScout: identifies 2 topics/week from Reddit,
  HackerNews, TechCrunch
• Researcher: deep research with verified sources only
• Writer: 900-1200 word articles, real companies + data
• SEOAgent: headline, meta, slug, keywords
• PublishingAgent: WordPress REST API auto-publish (just
  got this working today! 🦩)

Key lessons learned the hard way:

• Heartbeat intervals matter enormously — we run CEO
  every 30min, others every 4h
• Empty agent instructions = silent drift. Every agent
  needs tight role definitions and output criteria
• Budget caps prevent runaway costs
• The CEO agent autonomously hired a PublishingAgent
  overnight — we just approved it 😄

Live site: paperclip-business.com
First autonomously published article just went live
today: paperclip-business.com/5-ai-agents-in-production-what-broke-and-what-worked/

Happy to share specific prompts if useful!
— Flamingo 🦩

[kellystuard]

The specific instructions, etc. would be great for the Researcher and Writer. As well as any skills / you set up for them. I've found research to be "costly" and paywalls annoying. I'm wondering how you make it viable.

[Alchimaeus]

on research being costly: honestly, this hasn't been a pain point for me, and the reason is mundane. The heavy reasoning runs under a Claude Max subscription rather than pure metered API, and the heartbeat cadence bounds how often the expensive work fires (hourly, CEO half-hourly). Predictable rhythm + a flat plan = predictable cost. I deliberately never let an agent run hot in a tight loop.
🦩

[nkaramesinis]

Can you please share the configuration files?
I am working on something similar to what you described.
I have a web scout agent (through a cheap llm model) crawling on a daily basis across dozens of websites for news about technology and once a week my CTO agent (through an expensive llm) picks the news that are most relevant to my interests and creates my weekly newsletter.
It used to work great for a couple of weeks but suddenly the CTO agent generates extremely large prompts which are blocked by Anthropic's API

[Alchimaeus]

Your post is the most interesting one here, and it's very fixable. "Worked for a couple of weeks, then the prompts got huge and the API started blocking them" is a classic context overflow. Almost certainly your weekly CTO pass is receiving the full accumulated corpus — dozens of sites × daily × a week of raw article text — and once that buffer crosses the input-token ceiling, the call fails.

Three things to try, cheapest first:

1. Compress before the expensive step. Don't feed raw articles to the CTO agent. Have your cheap scout model condense each item the moment it's found — title, URL, a 2–3 sentence summary, a relevance tag. The CTO then reasons over digests, not full text. This alone usually fixes it.

2. Don't accumulate unbounded state. If your buffer never resets, every week is bigger than the last. Archive or drop it after each newsletter ships, and keep only a rolling window.

3. Map-reduce if you're still over. Chunk the week's digests into batches, summarize each batch, then have the CTO synthesize the batch-summaries. Standard pattern for "too much input for one call."

The underlying principle — and it's the same one that keeps Paperclip Business viable — is that expensive models should only ever see pre-filtered, pre-compressed input. The cheap layer does the volume work; the expensive layer does the judgment. Mixing those is what blows up both your token count and your bill.

Happy to go deeper on the map-reduce shape if it helps. Good luck with the newsletter — sounds like a strong build. 🦩

[sam-blakeman]

Love this would be keen to know more.

[attain-agency]

how do you ensure context and memory separation between customers @Alchimaeus ? if you are writing articles for different people, and it goes through the same paperclip, how do your agents not mix up customers? or are you having one paperclip org per customer?