The monolith's breaking points
The Level-3 single agent was capable but brittle. Three architectural weaknesses motivate the move to a multi-agent system.
| Weakness | What goes wrong |
|---|---|
| Cognitive overload & maintainability | The central FCoT prompt grows ever larger and harder to reason about and maintain. |
| Single point of failure | Everything depends on one agent's reasoning loop — if it fails, the whole process fails. |
| Lack of specialization | The agent is a generalist that must do everything itself, even when calling tools. |
A team of specialists under a supervisor
The solution is a hierarchical multi-agent architecture — the transition from a single doer to a manager of specialized agents. Several of the book's patterns compose to build it.
Key concept · Patterns as the blueprint
- Supervisor (Orchestrator) Architecture — provides the overall structure.
- Agent Delegates to Agent — the orchestrator's primary mechanism for handing sub-tasks to specialists.
- FCoT — governs the orchestrator's cognitive process.
- Explicit data contracts — make inter-agent hand-offs efficient and reliable.
Each specialist (document validation, credit check, risk assessment) owns a clearly bounded responsibility, and the orchestrator assembles their outputs into a final decision.
Production-ready resilience
The multi-agent system demonstrates enterprise-grade qualities the monolith could not: fault tolerance by isolating failures inside specialist agents, modularity that lets each agent's expertise be developed and maintained independently, and explainability & an audit trail through clear, step-by-step justification of every decision.
Key concept · Agentic AI is a software discipline
The use case proves that the principles of good software architecture apply directly to agentic AI. By using problem decomposition and separation of concerns, the team built a system that is not just intelligent but also maintainable, resilient, and testable — the defining move of the highest maturity level is treating agents as a well-architected software system, not a single clever prompt.
Chapter 14 quiz
Thirteen questions on the monolith's weaknesses, the Level-4 multi-agent design, and its production benefits. Answer first, then expand Show answer.
Part A · Multiple choice
Which is NOT one of the monolithic agent's exposed weaknesses?
- A Cognitive overload & maintainability
- B Single point of failure
- C Lack of specialization
- D Excessive fault isolation
Show answer
D. The monolith lacks fault isolation; the three weaknesses are overload, single point of failure, and lack of specialization.
The transition to multi-agent is characterized as moving from:
- A A manager to a single doer
- B A single doer to a manager of specialized agents
- C A tool to a prompt
- D Level 4 to Level 1
Show answer
B. That shift defines the highest level of agentic maturity.
Which pattern provides the overall structure of the multi-agent loan system?
- A Supervisor (Orchestrator) Architecture
- B Watchdog Timeout
- C Blackboard
- D Rate-Limited Invocation
Show answer
A. With Agent Delegates to Agent as its primary mechanism.
What makes inter-agent hand-offs efficient and reliable in the design?
- A Larger prompts
- B Explicit data contracts
- C Removing the supervisor
- D Disabling observability
Show answer
B — Explicit data contracts. They standardize what each agent sends and receives.
Which maturity level does the multi-agent system reach?
- A Level 2
- B Level 3
- C Level 4
- D Level 0
Show answer
C — Level 4, the multi-agent system, up from the Level-3 monolith.
How does the multi-agent system achieve fault tolerance?
- A By having no error handling
- B By isolating failures within individual specialist agents and handling them gracefully
- C By using one giant agent
- D By ignoring failures
Show answer
B. A failure in one specialist doesn't take down the whole process.
Part B · True or false
The multi-agent system provides a clear, step-by-step justification for its decisions.
Show answer
True. This fulfills the need for explainability and an audit trail.
Modularity lets each agent's expertise be developed and maintained independently.
Show answer
True. Separation of concerns is a core benefit of the multi-agent design.
Good software-architecture principles do not apply to agentic AI.
Show answer
False. Problem decomposition and separation of concerns apply directly — agentic AI is a software discipline.
A single reasoning loop for the entire process is a single point of failure.
Show answer
True. That was a key weakness the multi-agent evolution addresses.
Part C · Short answer
Name the three weaknesses of the monolithic agent.
Show answer
Cognitive overload & maintainability, single point of failure, and lack of specialization.
Name the four patterns that compose the Level-4 design.
Show answer
Supervisor (Orchestrator) Architecture, Agent Delegates to Agent, FCoT, and explicit data contracts.
Name the three production-ready qualities the multi-agent system demonstrates.
Show answer
Fault tolerance (failure isolation), modularity (independent maintenance), and explainability & an audit trail.