Stumbling into AI: Part 5–Agents (rmoff.net)

This entry in a short learning series digs into “Agents” — software that autonomously perceives, decides and acts — with a focus on how LLMs are shaping modern agentic systems. The author surveys definitions (Wikipedia’s software/intelligent agent, LangChain’s LLM-driven control-flow view, Simon Willison’s tool-loop definition), contrasts agents with static workflows, and gives concrete examples (from a simple “order biscuits” loop to coding assistants and travel-booking demos). The post emphasizes that today’s enterprise agents are usually limited in autonomy — often hybrid systems mixing LLM reasoning with rule-based IF…ELSE logic — and that industry discussion still wrestles with vocabulary (Agents vs Agentic AI). On the engineering side the write-up is practical: an agent can be built from an LLM API, tool/integration endpoints (MCP, RAG, external services), and control logic, but robust systems need memory, planning, reflection, validation, observability, guardrails and human-in-the-loop controls. Key implications for AI/ML practitioners are clear: memory/context and tool orchestration materially improve output, hybrid architectures balance autonomy and safety, and multi-agent systems decompose responsibilities for scalability. The note is framed by the author’s interest in Apache Flink and Confluent streaming agents and points readers to design-pattern resources and talks for deeper technical guidance.