Select appropriate architectural patterns (workflow, agentic, augmented LLM)
Choose between an augmented single call, a predefined workflow (chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) and an autonomous agent based on how predictable the path is, and justify the choice by cost, latency and auditability.
Key points
- 1
Anthropic's guidance: find the simplest solution possible and only add complexity when it measurably improves outcomes. Many applications need only a single augmented LLM call (retrieval, tools, memory) with a structured output.
- 2
Workflows orchestrate LLM calls through predefined code paths; agents let the model dynamically direct its own process and tool use. Workflows give predictability, reproducibility and bounded cost; agents give flexibility at higher cost and latency.
- 3
Prompt chaining: fixed sequential steps with programmatic gates between them; use when a task decomposes into known subtasks and a check between steps catches errors.
- 4
Routing: classify the input and send it to a specialised prompt and tool set; use when inputs fall into distinct, known categories handled differently (six email types, four resolution procedures).
- 5
Parallelization: sectioning splits independent subtasks and runs them at once; voting runs the same task several times for confidence. Use when subtasks are independent or multiple perspectives are needed.
- 6
Orchestrator-workers: a central LLM decomposes the task dynamically, delegates to workers and synthesises; use when the subtasks cannot be predicted in advance.
- 7
Evaluator-optimizer: a generator and an evaluator loop against explicit criteria; use when clear evaluation criteria exist and iterative refinement measurably helps (brand-rule compliance). Bound the loop with a round limit.
- 8
Agents fit open-ended problems where the number and order of steps depend on tool results and the environment gives verifiable feedback (tests, checks). Bound them with turn limits, budgets and human checkpoints.
- 9
A predictable exception (a clarifying question in 15% of cases) is a branch in a workflow, not a reason to switch to an agent. Hybrid designs are normal: route by predictability, workflows for the known 80%, a bounded agent for the open-ended 20%.
- 10
Justify pattern choice in terms stakeholders care about: workflows are auditable and cheap per run; agents cost more tokens and vary per run; evaluator loops and parallel workers add latency.
- 11
Distractor patterns: "make it an agent for flexibility" when the path is known; orchestrator-workers for a handful of dependent steps; evaluator loops on an interactive button that must feel instant; a single agent for everything when audit needs reproducible paths.
Read the source
Test yourself on Select appropriate architectural patterns (workflow, agentic, augmented LLM)
Ten questions, with the answer and explanation after each one.