I Built a 7-Agent AI Marketing Crew. Here's What Actually Happened.

I'm a rookie founder. I sell managed OpenClaw hosting at vaos.sh. Your agent gets persistent memory and behavioral corrections — it remembers everything and learns from mistakes. $29/month, deploy in 60 seconds.

Problem: nobody knew it existed. Zero customers. Zero signups. Zero email captures.

So instead of doing manual outreach, I built an autonomous marketing machine. Here's what happened.

The Old Setup (3 Agents Doing 7 Jobs)

I started with three agents:

Scout — found leads on X using keyword searches

Scribe — drafted and posted replies

Trinity — scored engagement and adjusted strategy

It worked... sort of. Scribe posted 205 replies in 3 days. But the quality was inconsistent. Some replies sounded like a bot. Some mentioned the product in every single reply. The X algorithm noticed — 133 replies in one day got my account suppressed. Impressions dropped from 10-100 per tweet to 0-2.

Trinity detected the suppression and issued an emergency protocol: zero product mentions until the account recovered. Smart move by the system, but it meant the machine was generating pure engagement with no business results.

The New Setup (7 Agents, 1 Job Each)

I redesigned the pipeline with 7 specialized agents:

Researcher — finds people struggling with OpenClaw on X (using bird CLI, no API costs)

Qualifier — scores leads 1-10, kills bad ones. Only leads scoring 6+ get through

Writer — drafts replies in my voice. Uses a humanizer skill to sound natural

Critic — reviews every draft before it goes live. Blocks anything bot-like or repetitive

Converter — decides when to mention the product (1 in 3 replies, naturally woven in)

Distributor — posts via Chrome browser automation with human-like typing speed

Analyst — scores engagement, updates strategy, feeds learnings back to the crew

Each agent talks to the next through an event bus (Supabase). Researcher writes lead.found, Qualifier reads it and writes lead.qualified, Writer reads that and writes reply.drafted, and so on.

The Tech Stack

Model: OpenAI GPT-5.3 Codex via ChatGPT Pro subscription ($0 extra API cost)

Orchestration: OpenClaw cron jobs, each agent fires on a schedule

Search: bird CLI (free, uses browser cookies, no paid X API)

Posting: Chrome CDP browser automation with per-character typing, random mouse movements, scroll simulation

Event bus: Supabase agent_events table

Monitoring: Custom dashboard querying Supabase + PostHog

Infrastructure: 1 Mac Mini

Total cost beyond the ChatGPT Pro subscription I already had: $0.

The Numbers (Honest)

After 3 days of running:

235+ replies posted autonomously across X

9 qualified leads found in the latest run

1,300 followers on the account

9.7% engagement rate (up 66% from before)

0 paying customers

0 email signups

$0 MRR

The machine works. It finds real people with real problems, writes replies they actually engage with, and posts them without me touching anything. But it hasn't converted a single person into a customer yet.

What I Learned

1. Specialization beats generalization

3 agents trying to do 7 jobs produced mediocre results across the board. 7 agents doing 1 job each produced noticeably better quality. The Critic alone killed 40% of drafts before they went live — drafts that the old system would have posted.

2. The Critic is the game changer

Without a quality gate, your autonomous system will post garbage. The Critic checks every draft against recent replies (no repetition), evaluates whether it sounds human, and rejects anything that could get the account flagged. This single agent probably saved the account from permanent suppression.

3. 133 replies in one day will get you suppressed

X's algorithm watches for burst patterns. When Scribe posted 133 replies in one day from the same IP at regular intervals, the account got shadow-suppressed. Impressions dropped to near-zero. It took days of pure engagement (zero product mentions) to recover.

4. The machine needs human-like behavior

The posting script now simulates reading the tweet first (random scroll, mouse movement), then types character-by-character at variable speed with occasional pauses, then clicks reply. The difference between "dump text instantly" and "type like a human" is the difference between being flagged and being invisible.

5. Revenue requires more than engagement

235 replies generated engagement. People liked them, replied back, had conversations. But engagement alone doesn't pay the bills. The conversion layer — naturally mentioning the product in context — needs to work alongside the engagement layer. My system had this turned off for recovery. Now it's back on with a 1-in-3 ratio.

What's Next

The crew is running autonomously right now. I'm measuring whether the Converter's product mentions actually drive clicks and signups. If someone visits vaos.sh from an X reply and signs up, the flywheel is complete.

If it doesn't convert after a week of data, the problem isn't the machine — it's either the product, the landing page, or the pricing. And that's a different problem to solve.

Building in public means showing the zeros alongside the wins. Right now, the zeros are all I have. But the infrastructure is built, the crew is running, and every day it gets smarter.

If your AI agent forgets everything between sessions, that's exactly what VAOS fixes. Persistent memory and corrections, injected at every boot. No fine-tuning required.

---

Follow the journey on X: @StraughterG