The shared building block: the augmented LLM

Every pattern on this site is built from one block: a large language model augmented with retrieval, tools, and memory. Anthropic describes a model that can generate its own search queries, select the right tools, and decide what to remember.

Google's whitepaper frames the same idea as three layers — the model, an orchestration layer that reasons (via ReAct / Chain-of-Thought / Tree-of-Thoughts), and a tools layer that reaches the outside world. Different words, same anatomy.

Workflows are systems where LLMs and tools are orchestrated through predefined code paths. Agents are systems where LLMs dynamically direct their own processes and tool usage. — Anthropic

Workflow vs Agent: the real dividing line

The dividing line is who controls the flow. In a workflow, the control flow is fixed in code you wrote; the LLM fills in the steps. In an agent, the LLM itself decides what to do next and keeps control over how the task gets done.

OpenAI puts it bluntly: agents are systems that independently accomplish tasks on your behalf. Apps that embed an LLM but don't let it control execution — simple chatbots, single-turn calls, sentiment classifiers — are not agents.

What changes when you cross the line

  • Control flow: hand-written code paths → the model's own decisions.
  • Predictability: high and testable → lower, emergent at run time.
  • Cost & latency: roughly fixed → variable, usually higher.
  • Best when: steps are knowable up front → steps are open-ended.

When to use which — start simple

Anthropic's overarching advice: find the simplest solution possible, and only increase complexity when it demonstrably improves outcomes. Agentic systems trade latency and cost for better task performance — so spend that budget deliberately.

Use a workflow when the task decomposes into predictable steps. Reach for an agent on open-ended problems where you can't predict how many steps are needed or hardcode a fixed path.

OpenAI's three triggers for building an agent

  • Complex decision-making — nuanced judgment, exceptions, context-sensitive calls (e.g. refund approval).
  • Difficult-to-maintain rules — rule sets that have grown unwieldy and error-prone (e.g. vendor security reviews).
  • Heavy reliance on unstructured data — interpreting documents or conversing naturally (e.g. an insurance claim).
  • Otherwise: a deterministic solution may suffice.

Six? Five? On the 'how many patterns' question

The article that prompted this site talks about 'six patterns'. That count is a presentation choice, not a substantive disagreement — so it's worth pinning down.

Anthropic's authoritative taxonomy is: one building block (the augmented LLM) + five workflow patterns (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) + the autonomous agent. Popular write-ups land on 'six' either by counting the five workflows plus the agent, or by promoting tool-use / context-augmentation / reflection to a standalone pattern.