His observation, stated plainly: an agent without a team is just a prompt with no context. Nobody hires a writer with no research team, no analytics, no strategy. You don't hire individuals. You build teams.

The Problem With Single Agents

He tried OpenClaw. He tried standalone agents with tool use. He tried the "AI employee" playbook. They kept failing for the same reason.

The research agent didn't know what the content agent wrote last week. The analytics agent's findings didn't reach the strategy agent. Context lived in prompts and died in sessions. The coordination overhead ate the efficiency gains.

A single agent, no matter how capable, can't run a growth operation. Growth requires coordination across functions. That coordination used to require humans. It still requires coordination — but it doesn't require human coordination.

What He Built Instead

A swarm framework where agents operate as a specialized team:

  • Research agent: identifies opportunities, trends, competitor moves
  • Strategy agent: converts research into testable hypotheses
  • Content agent: produces content for the current hypothesis, informed by what worked before
  • Analytics agent: measures performance, feeds results back to research
  • Coordination layer: routes outputs between agents, maintains shared context

Each agent has its own role, its own tool access, its own context — and they share knowledge through a structured handoff protocol.

What Changed

With individual agents: 2-5 experiments per week (limited by coordination bottlenecks). With the swarm: velocity is no longer the constraint. The loop runs continuously. Research informs strategy which informs content which generates data which informs research.

The human's role: set direction, review edge cases, make calls that require judgment. Everything else runs.

The Transferable Principle

This pattern — teams of specialized agents over individual general agents — shows up repeatedly in production use cases. Claude Code Review dispatches a team. LangChain's GTM agent spawns one subagent per account. The architecture scales; individual agents don't.