← All courses
Advanced · 7 lessons
Multi-agent orchestration: LangGraph, CrewAI, and agent teams
When one agent is not enough: supervisor, swarm, and pipeline topologies, LangGraph state machines, CrewAI role-based crews, hand-offs and shared memory, parallel coding agents in git worktrees, run economics, and a capstone — a three-agent team.
- 01When one agent is not enoughThe limits of a single agent and the three base topologies — supervisor, swarm, and pipeline: how to choose an architecture and when multi-agent only hurts.40 min
- 02LangGraph: state machines for agentsStateGraph in practice: typed state with reducers, nodes and edges, conditional routing, the checkpointer, and human-in-the-loop via interrupt.55 min
- 03CrewAI: role-based agent teamsAgent, Task, and Crew: roles, goals, and backstories, sequential and hierarchical processes, delegation, crew memory, and the line beyond which you need LangGraph.50 min
- 04Hand-offs, shared memory, and conflictsControl transfer via Command and parallelism via Send, blackboard and message-passing shared-memory patterns, and conflict resolution through reducers and an arbiter.55 min
- 05Parallel coding agents and git worktreeSeveral coding agents over one repository: why a shared directory is a wreck, a worktree and branch per stream, a queue with a concurrency limit, and safe merges.50 min
- 06Cost, latency, and failure containmentThe economics of multi-agent runs: the ×15 token multiplier, cost math, parallelism versus chains, timeouts, retries, budgets, and degradation to a single agent.45 min
- 07Capstone: a three-agent teamBuilding the spec → code → review pipeline: the architect writes the spec, the coder implements, the reviewer sends back fixes — looping to green tests with an iteration limit and cost tracking.60 min