Building your AI workforce.
Specialized AI Employees vs general models. Orchestration patterns. When agents hand off to humans — and when they don't.
Tutor · 39sYour AI Tutor
Module 04 · Building your AI workforce.
A workforce isn't one giant brain. It's a hundred specialists who hand off work cleanly. Same is true for an AI workforce — and most companies get this exactly wrong.
By the end of this module
- Choose between general-purpose and specialist agents
- Pick the right orchestration pattern for your workflow
- Design human-in-the-loop checkpoints that don't slow you down
Specialist beats generalist
One AI Employee with one job, doing it perfectly, is better than a god-tier general agent trying to do everything. Specialists are easier to debug, cheaper to run, and replaceable when models improve. General agents are stuck in a 'just one more prompt' spiral forever.
01 · Specialist beats generalist
Tutor explanation in production
Three orchestration patterns
Sequential: agent A finishes, hands to B, who hands to C. Best for clear-input-clear-output processes. Parallel: agents A, B, C run simultaneously, results merged. Best for research and enrichment. Conditional: agent A decides which of B, C, D to call. Best for triage and routing. Most real workflows mix all three.
02 · Three orchestration patterns
Tutor explanation in production
Human in the loop, but not in the way
Three places humans add value: (1) approving high-stakes outputs (contracts, public-facing content), (2) handling exceptions the agents flagged, (3) reviewing performance weekly. Humans should not be in the loop for routine cases — that defeats the point.
03 · Human in the loop, but not in the way
Tutor explanation in production
Do this · before the next module
Inventory your specialists.
List the agents you'd need for your first pipeline. Give each a one-sentence job description. If two agents have overlapping jobs, merge them or split them more.
Sketch the orchestration.
Draw boxes for each agent. Draw arrows for handoffs. Mark where a human approves. The picture should fit on one page.
Define exception handling.
What happens when an agent fails? What happens when output looks weird? Who gets paged? This is the difference between a system you trust and a system you babysit.
Workbook · 5-minute exercise
Sketch your first pipeline on a single sheet of paper. Boxes for agents, arrows for handoffs, stars where humans approve. Take a photo.