Descriptions:
Ben AI demonstrates how to build and run parallel sub-agent workflows inside Claude Cowork, using a 150-lead sales pipeline as the primary example: 15 sub-agents qualify all leads simultaneously against ICP criteria (SEO-focused marketing agencies in the US), a second wave of 18 agents enriches the 82 qualified leads with additional research, and a third batch writes personalized outreach messages—completing the full three-stage pipeline in approximately two minutes.
The video explains the underlying architecture of Claude Cowork’s agent system: skills store task-specific instructions, commands chain multiple skills into automated sequences, and the Agents tab within a plugin saves sub-agent role definitions for reuse. A key practical point is that users must explicitly instruct Claude to spin up parallel sub-agents in their prompt—without that instruction, Claude will process items sequentially. Ben walks through setting this up from scratch: prompting for lead qualification with a specified sub-agent count, saving the output as a skill, defining agent behavior files, and combining everything into a single slash command that triggers the full workflow.
The video also covers meaningful limitations: sub-agent workflows perform best on batches of 100–200 items, larger volumes are better served by dedicated automation platforms like n8n or Make.com, and heavy sub-agent usage requires a Claude Max subscription due to token consumption. Ben distinguishes sub-agent patterns (parallel isolated tasks) from multi-agent team setups (agents collaborating on a shared objective), referencing Anthropic’s own use of multi-agent teams to build a C compiler in two weeks. The result is a practical, replicable blueprint for anyone automating lead research, content generation, or data enrichment at scale.
📺 Source: Ben AI · Published March 04, 2026
🏷️ Format: Hands On Build
