Skip to content

Module 10 — Agentic AI

Level: Intermediate to Advanced · Time: ~6 hours including labs · Exam relevance: CCAR-P D1 (primary), D5

An "agent" is one of the most overused words in AI. This module gives you a precise vocabulary for it, shows you the simpler designs that solve most problems, and then teaches you how to build, supervise and harden real agents with Claude. The thread running through every page is the same piece of advice Anthropic's engineers give: start with the simplest design that works, and add autonomy only when it earns its cost.

Learning objectives

By the end of this module you can:

  1. Distinguish a single augmented LLM call, a workflow (code decides the path) and an agent (the model decides the path), and argue when not to build an agent.
  2. Recognise and implement the five common workflow patterns: prompt chaining, routing, parallelisation, orchestrator–workers and evaluator–optimiser.
  3. Write a correct agent loop against the Messages API, including every stop_reason and parallel tool results.
  4. Design a multi-agent system with an orchestrator and subagents, and explain its context isolation benefit, its token-cost multiplication and how failures propagate.
  5. Place human-in-the-loop controls: approval gates, confidence thresholds and escalation paths.
  6. Choose between a manual loop, the SDK Tool Runner, the Claude Agent SDK and Claude Managed Agents.
  7. Make an agent reliable with budgets, stopping conditions, idempotent tools, checkpoints and observability.

Page order

# Page What you learn
1 Workflows vs agents The complexity ladder and the "should this be an agent?" test
2 Workflow patterns Augmented LLM plus five composable patterns, with diagrams and code
3 The agent loop The observe–decide–act cycle in code, and every way it ends
4 Multi-agent architecture Orchestrator and subagents, isolation, cost and failure propagation
5 Human in the loop Approval gates, confidence thresholds and escalation
6 Managed Agents, Agent SDK, Tool Runner or manual loop The four ways to build an agent on Claude
7 Agent reliability Budgets, stopping conditions, idempotency, checkpoints, observability

Prerequisites

Exam relevance

This module underpins the following Claude Certified Architect – Professional (CCAR-P) objectives. Objective IDs and wording come from registry.yaml.

Objective Summary Pages
CCAR-P-1.3 Select architectural patterns (workflow, agentic, augmented LLM) 01, 02, 06
CCAR-P-1.4 Design multi-agent systems and orchestration strategies 03, 04
CCAR-P-1.5 Apply decomposition techniques to complex problems 02, 04
CCAR-P-5.3 Apply human-in-the-loop validation 05
CCAR-P-5.2 (secondary) Identify risks, limitations and failure modes of LLM systems 04, 07
CCAR-P-3.1 (secondary) Evaluate tool and agent configuration for capability bloat 03, 06
CCAR-P-4.6 (secondary) Monitor performance with logging and observability tooling 07

The CCAR-P study guides for Domain 1 and Domain 5 link back here.

Practise it

Next module

Continue to Module 11 — RAG and knowledge systems, where retrieval becomes one of the tools an agent can call. Agent security (excessive agency, tool misuse, agent identity) is covered in depth in Module 12 — AI security.