What an AI agent actually is (and isn't)

"Agent" is the most overloaded word in AI. Vendors slap it on any LLM that calls a tool. To build reliable systems you need a sharper definition, and the most useful one draws a single line — who controls the path?

Anthropic's framing, which has become the industry's shared vocabulary, splits agentic systems into two kinds:

• Workflows — systems where LLMs and tools are orchestrated through predefined code paths. You write the control flow; the model fills in steps. Predictable, testable, cheaper.
• Agents — systems where the LLM dynamically directs its own process, deciding what to do next and when it's done. The model controls the path. Flexible, but costlier and harder to make reliable.

Most "agents" you see demoed are actually workflows, and that's fine — often it's the right choice. The distinction isn't about prestige; it's about knowing which reliability problems you've signed up for.

Underneath both sits the real building block: the augmented LLM — a model equipped with retrieval, tools, and memory, able to generate its own queries, pick tools, and decide what to keep. Learn this atom first; every pattern is just augmented LLMs arranged differently.

The most important, most ignored advice in this whole field: often you shouldn't build an agent at all. Anthropic's guidance is to find the simplest solution and add complexity only when it demonstrably helps — which "might mean not building agentic systems at all." A single well-prompted LLM call with good retrieval and a few examples handles a huge share of tasks. Reach for a workflow when the steps are predictable. Reach for a true agent only when you genuinely can't predict the number of steps or hardcode the path, you can verify progress from real feedback, and the environment is trusted enough to hand the model the wheel.

Why be this disciplined up front? Because agents "trade latency and cost for better task performance" and introduce "the potential for compounding errors." Those aren't edge cases — they're the central reliability challenges this course exists to address. Start simple; earn each increment of autonomy.