Prompt engineering designs instructions that help an LLM respond accurately and in the desired format. For an agent, the main design surface is the system prompt: developer-defined instructions that persist across turns and set its permissions, constraints, and tool-use policies. User prompts change each turn and are outside the developer’s control. section 1.5.2 section 2.3
An agent should work on the user’s behalf, not merely converse. Translate that goal into instructions for when to act, when to clarify, and how to use available tools. section 2.3.2
Act when the request is clear
A weather request should lead an agent with a search tool to search, rather than ask whether it should search. A system prompt can state: “When the user’s intent is clear, execute immediately without confirmation. Only when intent is unclear, ask minimal questions to clarify.” section 2.3.2
A static prompt is only part of an agent’s input. Search results, code-execution results, and database results can enter the context as work proceeds. Context engineering extends prompt design by deciding what information to retain and provide, when, and in what form. section 1.5.2
