Skip to main content

Module 7: Agents, Sub-Agents and AI Teams

Learning objectives​

By the end of this module, you will be able to:

  • Describe the agent loop (observe, plan, act, check) and name the four components every coding agent is built from: model, tools, memory, stopping rule.
  • Explain the three reasons a single agent stops scaling - context pollution, serial execution, one set of permissions - and recognise each one in a session.
  • Define a sub-agent, write a task contract for one, and say what the orchestrator does with the result.
  • Assemble a team of specialised agents (planner, implementers, reviewer, integrator) for a feature, using the delivery loop from Module 1 and the specs from Module 2.
  • Name the five failure modes of multi-agent systems and the control for each.
  • Keep a human in charge: where the checkpoints go, what each role is allowed to do, how cost is bounded.

Module content​

Chapter 1: Anatomy of an agent​

The loop, the four components, and what distinguishes an agent from a chat. Why a long single session degrades, with the symptoms you will see.

Chapter 2: Sub-agents and delegation​

Context isolation, specialisation, parallelism. The orchestrator / worker pattern. How Claude Code, Cursor and GitHub Copilot each implement delegation. The task contract, the handoff and the merge.

Chapter 3: Building an AI team​

Roles, topologies (pipeline, fan-out, critic), a full worked example delivering a feature with five agents and one human, permissions per role, cost budgets, and the five failure modes.

Quiz​

Six questions.

Where this module sits​

Module 3 introduced agent mode as one of three ways to work with an assistant. This module goes inside it: what the agent is doing while you wait, why it eventually loses the thread, and how splitting the work across several agents - each with its own context, tools and permissions - gets past that limit without giving up control. Everything here assumes the specification and verification discipline of Modules 1 and 2; a team of agents without a spec is five ways to be wrong at once.