Multi-Agent Orchestration

By the end

What you'll build

• Decide when a single agent is the right answer and when a team genuinely earns its cost.
• Design a supervisor/worker system with narrowly scoped roles and tools.
• Implement handoffs using both message-passing and a shared blackboard, with documented contracts.
• Run a real two-role agent pipeline in the browser, graded deterministically on final shared state.
• Model the token, cost and latency multipliers a multi-agent team introduces.
• Diagnose multi-agent failures - misroute, corrupt success, deadlock, cascade - from a recorded trace.
• Add verifier/critic roles that catch bad handoffs before they reach the user.
• Map sandbox roles onto real-framework primitives (supervisor graphs, crews, handoffs).
• Ship a working two-agent research-then-draft pipeline on your own machine via an official quickstart.

Curriculum

What's inside

5 modules · 28 lessons

1. 01

   When One Agent Isn't Enough (and When It Still Is)

   6 lessons
   • Welcome: the honest truth about agent teams
   • The monolith first: why one agent usually wins
   • Three shapes: pipeline, supervisor/worker, network
   • Checkpoint: should this be a team at all?quiz
   • Lab: simulate a supervisor's routing tablepython lab
   • + 1 more lesson
2. 02

   The Supervisor / Worker Pattern

   6 lessons
   • Anatomy of a supervisor
   • Designing worker roles: specialisation & tool scoping
   • Agent lab: supervisor routes a ticket to the right workeragent lab
   • Fill in: the supervisor loopfill blank
   • Deterministic vs LLM routing
   • + 1 more lesson
3. 03

   Handoffs & Shared State

   6 lessons
   • Handoffs: message-passing vs shared blackboard
   • Agent lab: researcher -> writer through a shared blackboardagent lab
   • Lab: ordered blackboard writes (no races)python lab
   • Context engineering across a handoff
   • Flashcards: handoffs & shared stateflashcard sr
   • + 1 more lesson
4. 04

   Cost, Latency & Failure Modes of Agent Teams

   5 lessons
   • The multiplier problem: cost & latency
   • Lab: model the cost - single vs sequential vs parallelpython lab
   • Error analysis: read a failing multi-agent runtrace lab
   • Agent lab: add a checker role to catch a bad totalagent lab
   • Flashcards: failure modesflashcard sr
5. 05

   Capstone: A Two-Agent Research-then-Draft Pipeline

   5 lessons
   • Designing the two-agent pipeline: role contracts
   • Capstone lab: research-then-draft pipelineagent lab
   • Fill in: map sandbox roles to a real frameworkfill blank
   • Graduation: ship a real two-agent pipeline
   • Evidence check + course wrap-upquiz

The shape of it

How this course works