Agent Design Patterns: From Workflows to Autonomous Agents

Ask a candidate to design an "agent" for almost any task, and most will reach directly for the full observe-decide-act loop — the pattern agent-architectures-and-the-agentic-loop covers in depth — without pausing on the much larger space of options that sit below full autonomy. That's a real gap, because most systems that ship in production as "AI features" are not autonomous agents at all: they are fixed sequences of LLM calls, wired together in one of a small number of well-known shapes, that never once decide their own next step. Knowing the name and shape of each of those shapes — and knowing which one a given task actually calls for — is a distinct skill from knowing how the agentic loop works, and it's the skill this subject teaches.

This is deliberately the layer above the loop mechanics, not a restatement of them. agent-architectures-and-the-agentic-loop already owns the loop itself (observe → decide → act → observe), the ReAct pattern, plan-act-observe-reflect, single- vs. multi-agent at a glance, and the sharpest section in that subject — when not to build an agent, with the compounding-error arithmetic behind it. This subject assumes you've read that and starts one level up: given that full autonomy is a spectrum's far end, what does the rest of the spectrum look like, and what are the five named shapes that sit below it? It then covers three agent techniques that go beyond plain ReAct — Reflexion, ReWOO, and tree search for agents — each of which the sibling subject deliberately left out to keep the loop treatment focused. Read both together; neither one repeats the other.