agent-product-patternslisted

# Agent Product Patterns — workflow vs agent, and the building blocks The most expensive mistake in AI product design is building an **autonomous agent** when a **workflow** would have been cheaper, faster, more predictable, and easier to debug. Anthropic's central finding from production deployments: the most successful implementations use **simple, composable patterns — not complex frameworks or fully autonomous agents.** The PM's job here is to match the *minimum control mechanism* to the *observed failure mode*. Every step up the complexity ladder buys flexibility at the cost of latency, money, and debuggability — so you climb only when the rung below demonstrably fails. This skill gives you the vocabulary to make that call *before* the spec is written, so engineering builds the right shape the first time. ## 1. The core distinction: workflow vs agent Anthropic draws one line that governs everything else: - **Workflow** — an LLM (and tools) orchestrated through **predefined code paths.** The control flow is written by you, in code. The LLM fills in the cognitive steps, but *what happens next* is decided by your program. Predictable, consistent, testable, cheap to reason about. - **Agent** — an LLM that **dynamically directs its own process and tool use,** maintaining control over how it accomplishes a task. The control flow is decided by the model at runtime. Flexible, capable of open-ended work — and unpredictable, harder to test, more expensive, and cap