Skip to main content

CrewAI

Turn a single all-purpose agent into a small team where each member has a role, a goal and a specific task. This course teaches you when that trade is worth it, how to compose the team, and how to keep the bill and the debugging under control.

Course Duration: 5h​

What You'll Learn​

  • Decide, case by case, whether one agent, a crew or a plain script is the right answer
  • Define a CrewAI agent with a role, a goal, a backstory and its own model
  • Chain tasks with clear expected outputs and pass context from one to the next
  • Choose between a sequential process and a hierarchical process with a manager agent
  • Attach the right tools to each agent — read, search, save — and set their permissions
  • Enable delegation between agents and detect the loops it can trigger
  • Use short-term, long-term and entity memory without exploding the context window
  • Measure the number of calls, the cost and the latency of a run and compare with a single agent
  • Read a trace when a crew talks in circles and apply the concrete fixes that unblock it
  • Ship a 10-page product documentation written end to end by a four-agent crew

Prerequisites​

  • Course 30 — AI agents (tools, ReAct loop, guardrails)
  • Comfortable with Python packages, environment variables and a terminal

Course Modules​

  1. When several agents beat a single one
  2. Agents, roles and goals
  3. Tasks, dependencies and expected outputs
  4. Sequential and hierarchical processes
  5. Tools shared between agents
  6. Delegation and supervision
  7. Shared memory and context passing
  8. Cost, latency and limits of the approach
  9. Debugging a crew that does not converge
  10. Project: a documentation writing crew

The running project​

The same team of four agents runs through every module: an Analyst who reads a rough product specification, a Writer who drafts the sections, a Reviewer who checks consistency and terminology, and a Manager who arbitrates when the other three disagree. Each module adds one capability to that same crew — the four roles (2), chained tasks with typed outputs (3), a sequential run then a hierarchical one with the Manager as router (4), shared read and search tools (5), controlled delegation (6), a common memory that survives across tasks (7), a measured bill and latency (8), the fix for a crew that keeps handing the task around (9), and a final 10-page documentation delivered end to end (10). The model can be a hosted API or a local Ollama server — the choice is yours; the code changes by one line.

Assessment and certificate​

The course ends with a 40-question exam covering every module. On success, a certificate of completion is issued; its number is verifiable on the platform.

Free courses, by contrast, end with a 5-question quiz and a preview of the certificate, without certification.