CASE STUDY · MULTI-AGENT
Mojo: a 7-agent WhatsApp ops system.
A two-person creative team needed agency-scale output. Mojo is a 7-agent orchestration system running real client work, built in 36 hours.
The problem
A small creative team needed the output of a full agency. Creative sprints ran four to six weeks, and there was no headcount to scale them without breaking cost or quality.
The build
Seven agents orchestrated through WhatsApp: Ops for routing and Gmail drafting, Scout for research, Strategist for campaign angles, Writer for scripts and copy, Babel for localization, Producer for production briefs, and Mojo as the conductor.
It runs on the OpenClaw runtime with a custom TypeScript plugin, compressing weeks of work into days across 175+ languages on infra under $40 a month.
What broke
The first version used an LLM router and failed 50 to 70% of the time. A deterministic state machine with a fixed sequence and three approval gates replaced it, and got to 20+ clean end-to-end runs.
Hallucinated tools were fixed with tool-awareness injection, a double-dispatch race with in-memory locks, and self-approval with three-layer verification guards.
What it taught
The working system was roughly 10% LLM and 90% plumbing. Reliability came from constraining the agents, not from a smarter prompt.
Want the full build story?
The complete write-up covers the architecture, what broke, and what came next.