The moment I decided to scrap the OpenClaw setup wasn’t a dramatic failure. It was a spreadsheet.

I was auditing my Anthropic API costs trying to figure out why a simple content pipeline - research a topic, write a post, review it, queue it - was costing me $330/month. That’s $10 a day for one LinkedIn post.

When I broke it down by agent and task type, the answer was uncomfortable: most of the tokens weren’t going to the creative work. They were going to the orchestration.

---

What “AI Orchestration” Actually Costs You

Every time an OpenClaw agent runs, it loads its full context. Instructions, memory, tool definitions - and the logic of the workflow it’s supposed to execute.

That workflow logic looks like this: “Check if there’s a research proposal in the queue. If yes, pick it up and write a brief. If no, go back to sleep. Log your status. Move the state forward. Handle edge cases.”

That’s not creative work. That’s an if-statement. And I was paying language model rates to process it dozens of times a day.

Here’s the realization that reframed everything for me:

Real workflows can almost always be broken down into a set of determined steps. When that’s the case - it should be the infrastructure’s job to handle the steps, and the agent’s job to handle the creativity.

An AI agent deciding whether a research proposal is in the queue is waste. An AI agent synthesizing that research into a sharp brief - that’s value.

The problem with OpenClaw (and agent-native frameworks generally, for this type of work) is that they treat both things the same way. Everything runs through the model. Everything costs tokens. The architecture doesn’t distinguish between “work that requires intelligence” and “work that requires a cron job.”

---

What I Actually Rebuilt

The n8n version has no orchestrator.

Instead, the workflow is a state machine. Each of the four former agents - researcher, writer, reviewer, and the publisher queue - is now its own n8n workflow. A cron checks if there’s work to do at that stage of the pipeline. If there is, it does it, updates the pipeline state, and stops. No AI involved in that decision.

The AI only touches four things now:

1. Web research and source evaluation
2. Distilling topics and briefs from that research
3. Writing posts from briefs
4. Peer review against defined quality criteria
5. A simplified chat agent interface for pipeline briefs and human approvals

That’s it. No AI deciding when to run. No AI managing state. No AI handling tool errors. The workflow handles all of that. When a tool fails, n8n catches it and retries or alerts me - not because I wrote an AGENT.md instruction telling it to, but because that’s what workflows do.

---

What I Lost (And Got Back)

I genuinely liked talking to Oracle. Asking “what’s in the pipeline?” in plain English and getting a coherent briefing back - that’s a better experience than a dashboard. It’s faster, more natural, and it keeps you in the flow of the work.

I missed it enough that I built it back in - just scoped correctly this time.

Instead of a full orchestrator making decisions about the workflow, I added a simple tool-enabled chat agent - still through Telegram, still conversational, just scoped correctly. It can pull pipeline state, surface what needs review, and accept approvals and revision notes. That’s it. It has no control over when workflows run, no ability to manage state directly, no role in the automation logic.

The original OpenClaw version felt like texting an employee who also happened to be running the entire operation. This is more like texting a coordinator who reads you the dashboard and passes your notes along.

The distinction matters. The conversational interface was never the problem - it was a genuinely good experience. The problem was the AI being responsible for the infrastructure underneath it. Pull those two things apart and you get the best of both.

---

The Cost Comparison

	OpenClaw	n8n Rebuild
Infrastructure	Included	$14/mo (self-hosted)
Token costs	~$316/mo	~$30-56/mo (est.)
Total	~$330/mo	~$44-70/mo
Daily debugging	Near-daily	Rare
Reliability	Inconsistent	Consistent

The n8n rebuild hasn’t been running long enough to report final numbers with confidence - that’s what Part 3 is for. But directionally, it’s looking like 70-80% cost reduction for the same creative output and better pipeline reliability.

The engagement results from the OpenClaw period hold as a baseline: 32% increase in engagement and 8% increase in impressions over the prior 28 days. The writing quality drove those numbers. The infrastructure just needs to stop getting in the way.

---

The Rule I’d Give Anyone Evaluating AI Agents for Automation

OpenClaw isn’t bad software. For genuinely unpredictable, open-ended tasks - personal research, ad hoc problem solving, tasks where you can’t predict the steps in advance - an agent-native approach makes sense. That’s what it’s good at.

But for business workflows that run on a schedule and follow a defined process? You’re paying AI rates for infrastructure work.

The test is simple: can you draw a flowchart of the process? If yes, the flowchart should be code - or n8n, or whatever structured automation tool fits your stack. Use AI for the nodes in that flowchart that genuinely require judgment.

Everything else is just an expensive cron job.

---

Part 3 will have real numbers from 30 days of the n8n rebuild running in production - cost breakdown, reliability comparison, and whether the creative quality held without the OpenClaw scaffolding. Check back soon.