Not every task needs an agent. Some just need a well-tooled LLM call or a fixed sequence of steps — and reaching for autonomy too early trades predictability for flexibility you didn't need, at a real cost in latency, token spend, and compounding errors. Agent architecture is the discipline of picking the right shape for the job: a single loop, a router, a team of specialists, or a strict workflow — and knowing why.
This module starts with the foundational distinction between workflows (predefined code paths) and agents (models that direct their own process), straight from the Anthropic team that coined the framing most of the industry now uses. It then surveys the landscape of agent types — reactive, deliberative, hybrid, learning, and hierarchical — before zooming into why teams reach for multiple agents at all: context limits, specialization, and the coordination patterns (supervisor, hierarchical, swarm) that keep a multi-agent system from collapsing into chaos.
Watch the lessons in order, then continue on to Building Agents.
AI Engineer co-founder Shawn "swyx" Wang opens the AI Engineer Summit by defining what actually makes a system "agentic" — goal-orientation, tool use, control flow, and delegated authority — and why architecture is the decision that follows.
IBM's Martin Keen breaks down reactive, deliberative, hybrid, learning, and hierarchical agents — the base architectural vocabulary every more advanced pattern builds on.
Anthropic's Barry Zhang draws the line between workflows and agents, argues against building agents for everything, and walks through the composable patterns — routing, orchestrator-workers, evaluator-optimizer — that hold up in production.