15 - Agent Patterns & Multi-Agent

How to compose model calls, tools and agents into systems: the workflow patterns most production systems are made of, the patterns that make a single agent safe and reliable, when multiple agents genuinely help (and the evidence on when they don't), and the engineering that makes multi-agent systems work. A lab puts the central claim - feedback must add information - to a measured test.

Learning objectives 7-9 hours (notes + lab)

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

Choose and combine workflow patterns (chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) for a task, with cost and latency estimates
Wrap a single agent in guardrails, gates, human handoffs, sub-agents and verification loops that use real feedback
Decide, citing evidence, whether a task warrants multiple agents, and pick a topology
Engineer multi-agent context, state, communication, budgets, tracing and evaluation
Measure whether a pattern helps on your task rather than assuming it does

Prerequisites

Agent Foundations - the loop, tools, memory, planning
MCP & A2A - for agents across boundaries

Where This Module Fits

flowchart LR
    W["01 Workflow<br/>patterns"] --> S["02 Single-agent<br/>patterns"]
    S --> M["03 Multi-agent<br/>architectures"]
    M --> E["04 Multi-agent<br/>engineering"]
    W -.-> LAB["🧪 Lab: patterns<br/>under measurement"]
    S -.-> LAB

    style W fill:#e8e2d9,stroke:#ccc4b8
    style S fill:#d8dfe8,stroke:#b0bac8
    style M fill:#e8e0d4,stroke:#c8b89a
    style E fill:#ddd8e4,stroke:#b8b0c8
    style LAB fill:#dde4dc,stroke:#b0c4b0

Frameworks that implement these patterns are compared in Agent Frameworks; running them in production is Production Agents.

Chapter Map

#	Chapter	You will learn	Time
1	Workflow Patterns	Chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer; combining; cost and latency	50 min
2	Single-Agent Patterns	The single-agent default, guardrails, gates and HITL, sub-agents as tools, verification loops	45 min
3	Multi-Agent Architectures	The evidence, topologies (orchestrator, hierarchical, handoffs, network, blackboard, debate), MAST failure modes	55 min
4	Multi-Agent Engineering	Delegation and artifacts, communication, state and ledgers, budgets, tracing and evaluation	55 min
5	Q&A Review Bank	40 questions across the module	45 min

Code Lab

Lab	What you build	Runs on
Patterns Under Measurement	Single attempt vs self-review vs test-feedback loop vs best-of-n on HumanEval problems, graded on hidden tests with confidence intervals	A local OpenAI-compatible model (laptop) or any hosted API; ~1 hour for the full run

Mini-Project

Take a task from your work that someone proposed a multi-agent system for, and settle it with measurements:

A test set of 30+ real inputs with an outcome check (state, tests, or a rubric validated against 20 human labels).
A strong single-agent baseline (good tools, clear prompt, a verification loop with external feedback).
The proposed multi-agent design, with precise delegation schemas, an artifact store and budgets.
A comparison of quality with confidence intervals, tokens, cost and latency per input - including a single-agent variant given the same token budget.
A one-page recommendation, citing your numbers and the evidence in this module.

Review

Q&A Review Bank - consolidated questions for this module
Module quiz - every Check Yourself question in this module, in course order

Previous: 14 - MCP & A2A · Next: 16 - Agent Frameworks

Section Appendix

Summary & Key Terms - a quick recap of this section and its essential vocabulary.