Claude Academy
Sign in

CCA-F Certification Prep · lesson 3 of 10

Domain 1 — Agentic Architecture & Orchestration (27%)

The largest domain. Tests whether you can pick the right architecture for a given problem.

Topics

1.1 Workflow vs Agent

From Anthropic's Building Effective Agents:

  • Workflow: "LLMs and tools are orchestrated through predefined code paths."
  • Agent: "LLMs dynamically direct their own processes and tool usage."

Default to workflow unless dynamism is required. "Agentic systems often trade latency and cost for better task performance."

Test your scenario: can you predetermine the sequence of steps? If yes → workflow. If you need open-ended reasoning across N+ steps with ground truth checks at each step → agent.

1.2 The five patterns (memorize trade-offs)

PatternWhenCostLatencyPredictability
Prompt chainingSteps known in advanceLowPredictableHigh
RoutingMultiple specialized handlersLowLow (1 classify + 1 handler)High
ParallelizationIndependent subtasks OR consensusMidLow (parallel)Mid
Orchestrator-workersDynamic decompositionHighVariableLow
Evaluator-optimizerQuality matters > speedHighHigh (iterative)Mid

See agentic-patterns for full details.

1.3 Multi-agent coordination

  • Orchestrator holds the plan and the goal. Larger model usually.
  • Workers are focused, smaller, often Haiku.
  • Provenance flows up — each worker tags results with source IDs.
  • Shared mutable state is an anti-pattern between workers; pass results up via orchestrator.

1.4 Task decomposition

  • Break by independence when parallelizing.
  • Break by dependency chain when sequencing.
  • Break by expertise when routing (different tool sets per worker).
  • Tag every subtask with: input contract, output contract, allowed tools.

1.5 Session state management

  • Application owns conversation state — the API is stateless.
  • State object pattern beats replay for long-running flows.
  • Subagents should not share state; each gets a self-contained brief.
  • Use IDs (request_id, conversation_id, task_id) to thread observability.

1.6 Human-in-the-loop

  • Confirmation gates before destructive ops.
  • Escalation paths when confidence is low or tools fail.
  • "I don't know" is a valid agent action — better than hallucinating.

1.7 Reliability primitives

  • Capped iterations (avoid infinite loops).
  • Retry policies for transient tool errors.
  • Fallback to a simpler model or hand-off to a human.
  • Audit log of every tool call.

1.8 The three agent principles (from the source)

  1. Simplicity — minimum viable pattern.
  2. Transparency — surface planning steps to the user.
  3. Tool documentation and testing — invest in the agent-computer interface (ACI). Anthropic's SWE-bench team "spent more time optimizing our tools than the overall prompt." Small tool-interface changes (e.g., requiring absolute paths) often beat prompt tweaks.

Common question shapes

  • "Given this multi-step workflow, which pattern is most appropriate?" → Match dynamism to pattern.
  • "Orchestrator made a bad plan. Most likely cause?" → Vague subtask briefs, missing tool descriptions, too many tools.
  • "Two subagents need to share data — how?" → They don't directly; the orchestrator integrates.
  • "How do you bound an agent's runaway behavior?" → Iteration cap + permission denies + human gate for destructive ops.

Cheatsheet

  • Simplest pattern that works. Don't reach for agents when a workflow suffices.
  • Specialize workers, generalize the orchestrator.
  • Stateless tools, state-aware app.
  • Confirm before destruction.

See also