Prompt engineering for agentic AI requires a fundamentally different approach than prompting chatbots. Key differences include managing context rot across long task sequences, designing system prompts at the right altitude (avoiding both over- and under-specification), and structuring tool descriptions with clear boundaries. The post covers four essential agent prompt components: system prompts, tools, few-shot examples, and context state management. It also explains reasoning architectures like Chain of Thought, ReAct (Thought→Action→Observation loop), and Reflexion (self-correction). Practical patterns include just-in-time context loading, outcome-based prompts over procedure lists, dynamic persona priming, and minimal shared context in multi-agent orchestrator-worker setups. Five common mistakes are identified, including too many tools, vague success criteria, and overloaded context windows.
