Notes: Direct AI Agent Decisions (Goals, State and System Prompts)

Section 5 explains how an agent responds when the next step is not fully prescribed. The imagined folding robot carries the learner from a changed input to a running record of the work, different levels of autonomy, goals, constraints, standing instructions, saved preferences, and conditions for acting or waiting.

Lecture	Understanding developed
What Happens When a Workflow Meets an Unexpected Input? (Edge Cases)	An edge case is an unusual input or condition, and the application needs a supported response that can carry out the correction.
How Does an AI Agent Track Progress? (State and Multi-Step Jobs)	State records where a multi-step job has reached, while the context window limits what the model can consider during one turn.
How Much Can an AI Agent Decide? (Autonomy and Permissions)	Autonomy describes how much deciding and acting happens without a person; tools, permissions and approval rules separately determine what can proceed.
How Does an AI Agent Choose Between 2 Good Answers? (Goals and Constraints)	A useful choice advances the goal while satisfying its constraints, and an impossible combination should lead to a request for a changed constraint.
What Is the Difference Between a Goal and an Instruction?	An instruction states the work to perform, while a goal states what the answer should help decide or produce. When the desired result guides choices during the job, the work is goal-directed.
What Does a System Prompt Tell an AI Agent?	A system prompt supplies builder-defined standing directions, while the application controls information, tools and executable actions.
How Do Your Preferences Change AI Results? (Persona and Personalisation)	A persona is a role requested for a response; personalisation carries retained preferences into relevant later work.
When Should an AI Agent Act or Wait? (Goals and Action Limits)	Workload, deadline, cost and consequence determine when an allowed action should happen, wait or require approval.