AI developers urged to adopt agentic workflows instead of single-prompt chatbots

Two recent analyses highlight the evolving best practices for working with large language models (LLMs). The first outlines five core lessons from prompt‑engineering guides, emphasizing that prompts do not travel between services, responses are unpredictable, and that prompts may need ongoing monitoring as models change.

The second argues that relying on a single prompt is a dead end. It advocates “agentic workflows” that break tasks into iterative steps—planning, execution, critique, and refinement—and describes three effective patterns: a reflection loop that critiques drafts, tool‑use loops (ReAct) that let the model call external functions, and multi‑agent delegation that distributes subtasks among specialized agents. It also warns that managing state across these loops is essential to avoid token bloat and loss of goal focus.