Learn · Agentic architecture & orchestration · CCAR-F

Workflows vs agents

Six patterns from one call to a free agent, and the rule for picking: if the steps are known, let code control the path.

Key ideas

• Start with a single well-prompted call; add structure only when it's needed.
• Prompt chaining runs fixed steps with code gates between them.
• Routing classifies the input and sends it to a specialised handler or a cheaper model.
• Parallelisation is sectioning (independent parts) or voting (same task, several runs).
• Orchestrator-workers decides subtasks at runtime; evaluator-optimizer loops generate, critique, revise.
• An autonomous agent directs its own tools and needs guardrails and stop conditions.
• Known, fixed steps: prefer a code-controlled workflow over a free agent.

Worked example: Picking the pattern for an invoice pipeline

Every invoice is extracted, validated against the PO, then summarised. The steps never change, so use a prompt chain with a code gate after extraction, not an agent that decides what to do next. If invoices came in three distinct types, add a router in front.

Exam traps

• Choosing an autonomous agent for a task whose steps are fixed and known. A workflow is more reliable.
• Confusing orchestrator-workers (subtasks decided at runtime) with parallelisation (parts known up front).
• Building multi-step structure before checking whether one well-prompted call already works.
• Giving an agent open-ended tools with no guardrails or stop conditions.

Official docs

• Building effective agents
• Agent SDK overview

Practise CCAR-F questions · Next lesson: Coordinators and subagents