If you’ve been anywhere near the AI space lately, you’ve heard about OpenClaw. Nvidia compared it to what GPT was for chatbots. It has 300,000+ users. The pitch is simple and genuinely compelling: a personal AI agent that handles real work while you sleep.

I fell for it. Deliberately.

A few weeks ago I built a 4-agent LinkedIn content marketing pod on OpenClaw. The goal was dual-purpose: produce content at scale while I ran an honest technical evaluation of whether this thing could actually automate a real business workflow.

Here’s what happened.

---

The Build

The design took about a day. Four agents:

• Oracle (Orchestrator) - manages the pipeline, coordinates between agents, surfaces issues
• Tank (Researcher) - scours the web, builds research proposals on trending topics
• Neo (Writer) - takes research briefs and drafts posts in my voice and style
• Smith (Reviewer) - checks drafts against quality criteria before they enter the publishing queue

Yes, they’re all Matrix characters. No, I don’t regret it. If you’re going to spend weeks arguing with an AI about why it hallucinated a research brief, you might as well be arguing with an Agent.

---

Under the Hood

The model choices were deliberate and varied across all four agents - I wasn’t going single-vendor.

Oracle (Orchestrator) started on Claude Opus. She was managing the pipeline, fielding my status questions over Telegram, and coordinating handoffs between agents. Opus felt justified at first. But as costs climbed, I downgraded her to Sonnet, then eventually Haiku. The orchestration quality held up fine at lower tiers - which, in hindsight, was a signal I ignored. If Haiku can do the job, that tells you something about the nature of the job.

Tank (Researcher) ran on Gemini 2.5 Flash Lite. High-volume work, heavy tool use, needs to be fast and cheap - Gemini Flash was the right call there. Not a creative model, but that wasn’t the ask.

Neo (Writer) ran on Claude Sonnet 4.6. Creative writing, voice matching, tone consistency. This is where the model quality actually mattered and the token cost was justified.

Smith (Reviewer) ran on Grok-4-mini. Intentionally a different model from a different company with a different training profile. The idea: if Neo writes something and an Anthropic model reviews it, you might get agreement through shared biases. Grok brings a genuinely different perspective - it’ll push back where Claude might wave something through.

For the database, I used PocketBase - simple to set up, ran locally alongside the agents. I wrote a custom pb-cli script so the agents could interact with it directly from the command line without needing a full API integration.

And the end of the pipeline: a script that checked the PocketBase queue for approved posts and pushed them to LinkedIn automatically. Full end-to-end automation, no human hands on publish.

It was a well-designed system. That’s what made the failure modes so frustrating.

---

From design spec to running in the wild: a few days. That part impressed me.

The experience of working with Oracle through Telegram felt surreal - in the best way. I’d ask it what was in the pipeline and get back a clear status briefing. When something was off, I could diagnose it conversationally. It described its own internal issues. It felt like briefing an employee who happened to never sleep.

For the first week or so, it felt like the future.

---

Where It Broke

The researcher was the bottleneck. Every time.

That agent had the most breadth of work in the pipeline - find topics, evaluate relevance, build research proposals, hand work forward. And that breadth is exactly where OpenClaw agents struggle.

The failure modes were specific and frustrating:

Silent failures. When a tool errored out, the researcher didn’t surface it or retry. It just continued - quietly failing on every heartbeat, producing nothing, reporting nothing. No alert. No self-correction. Just a stalled pipeline I’d discover hours later.

Workflow amnesia. It would occasionally complete a research piece and then forget to move it into the review queue. Content just sat there. No notification. No follow-up. Gone until I went looking for it.

Hallucination. In one case, the researcher skipped research entirely and fabricated the whole piece. Not a summary of bad sources - a completely made-up research proposal. Confident. Detailed. Wrong.

I rewrote its AGENT.md file more times than I can count. Added guardrails. Added explicit step-by-step instructions. Added error handling directives. It would improve for a day or two, then drift into new failure modes. New issues, new patches, new issues. A moving target with no obvious end.

---

The Results (Yes, There Were Some)

Despite the researcher’s reliability problems, the content that did make it through the pipeline performed. The past 28 days showed a 32% increase in engagement and an 8% increase in impressions compared to the prior period.

The writing quality held up - but only after significant upfront work training the writer agent to match my tone. Teaching an AI to sound like you is a project in itself. Once it was dialed in though, it stayed dialed in. That part worked.

So: the creative output was good. The pipeline that fed it was not.

---

The Cost Problem

$330 in Anthropic API tokens over 30 days.

For one post per day. Plus near-daily debugging sessions with the researcher.

For context - a low-tier social media manager runs $300-500/month. The math almost works, except the social media manager doesn’t need you to debug their tools every morning.

The engagement lift was real. The writing was good. But $330/month for a workflow that required daily babysitting wasn’t what I had in mind when I started thinking about automation.

I started asking why it cost so much - and the answer pointed toward a more fundamental problem.

---

Why OpenClaw Is Overhyped

Here’s my honest take, and I say this as someone who was genuinely excited about it:

For non-technical users, OpenClaw feels like magic. Watching an agent wake up, do real work, and report back - that’s a compelling demo. The conversational interface is genuinely novel.

For serious business workflows, the cost and reliability math doesn’t hold.

The token cost issue isn’t a prompt engineering problem. It’s structural. Every time an OpenClaw agent runs, it loads its full context - instructions, memory, tool definitions, and the logic of the workflow it’s executing. That workflow logic is being processed by a language model on every heartbeat. An LLM that charges by the token.

You’re paying AI rates for work that has nothing to do with intelligence. You’re paying it to remember “if there’s a research proposal ready, draft a brief - otherwise, go back to sleep.”

That’s not creative work. That’s an if-statement.

And as you add more posts, more channels, more workflows - costs scale almost linearly. The architecture doesn’t get more efficient. It just gets more expensive.

---

The security implications and long-term maintenance burden of agent-native workflows are a separate conversation - but they’re real considerations before you commit to building anything serious on this.

I started thinking about what this workflow would look like if I treated it as what it actually was: a deterministic process with a few creative nodes inside it.

That thinking led to a full rebuild. Part 2 covers the realization, the rebuild, and what a more structured approach actually looks like.