Autonomous Content Org on €35/Month: What Actually Runs
This is a case narrative, not a product pitch. I run a content organisation across multiple sites with AI agents doing the routine work. Infrastructure sits around ~€35/month. The point is what has to be true for that to stay reliable.
For the operating-model thesis, see Company OS. For the discipline definition, see what is agentic engineering.
What “autonomous content org” means here
Not “AI writes whatever and posts it.”
It means specialised agents with clear roles:
- Signal / research — reads sources and drafts briefs
- Copy — writes in a defined voice for a defined audience
- Routing / CMO-style coordination — decides what is in scope this week
- Publish path — opens the change (PR / CMS) and waits
Humans approve anything externally visible. That approval gate is the product, not a temporary crutch.
Evidence I will stand behind
From systems I operate today (order-of-magnitude, not a finance audit):
| Signal | Rough figure |
|---|---|
| Live projects in the studio | 10+ |
| Containers on the stack | 30+ |
| MCP integrations | 12+ |
| Monthly infrastructure | ~€35 |
Source of truth for those numbers: site profile facts / llms-full. Older writeups also cite ~€26–€35 depending on VPS size and whether API spend is mixed in — treat €35 infra as the current infrastructure band, with LLM APIs billed separately.
The loop that has to work every day
- Intent — which sites and topics are in play
- Context — briefs, prior posts, brand constraints
- Execution — agents draft and assemble
- Tools — git, CMS, analytics, Slack/Telegram via MCP / workflows
- Approval — human yes/no on publish
- Feedback — performance feeds the next brief
Skip step 5 and you get spam. Skip step 6 and you get volume without learning. That six-step shape is why I talk about Company OS vs frameworks vs orchestration instead of “which agent library.”
Stack pieces (roles, not brand loyalty)
- Orchestration — n8n for triggers, retries, handoffs (n8n hub)
- Agent coordination / approvals — Paperclip-style issue + approval flow
- Model routing — LiteLLM
- Tool access — MCP gateway with allowlists (production lessons)
- Hosting — Coolify on a Hetzner VPS
Detail dump: /toolstack and earlier field notes such as AI agents as co-workers.
What breaks (and how we catch it)
| Failure | Symptom | Mitigation |
|---|---|---|
| Silent tool error | Agent says “done”, nothing published | Explicit incomplete status + alerts |
| Over-permissioned MCP tool | Blast radius across systems | Per-agent allowlists; write tools gated |
| Schema drift | Empty fields, no-op “success” | Versioned tools + smoke tests |
| Approval fatigue | Humans rubber-stamp | Only gate externally visible actions |
| Cost spike | API bill, not VPS | Route cheap models for drafts; reserve strong models for critique |
What this is not
- Not proof that every company should fire writers
- Not a claim that quality matches senior editorial work on novel topics
- Not a SaaS “Company OS” you install tomorrow
It is proof that a solo operator can run routine multi-site content ops with a small infra bill — if orchestration, permissions, and approvals are designed on purpose.
If you want the same shape inside a company
Start with one workflow, one approval policy, and one owner. Hiring signals: when to hire a fractional agentic engineer. Stockholm / Europe: /agentic-engineer-stockholm. Discovery: /hire.