深习堂
课程价格指南博客

开放课程Skill DevelopmentAI Courses

AI Agents for Beginners

Understand how agents carry out tasks, then sketch a first design with clear goals, steps, and limits

开始学习
全部章节
10
Locale
English
课程类型
开放课程

Learn how AI agents work through customer follow-ups, meeting notes, and other everyday tasks. Decide when a task needs tools, source material, saved state, or human approval. Then sketch a minimal agent design and check it for gaps and risks.

适合人群

  • Beginners who have heard of AI agents but are unsure how they differ from ordinary chat
  • People who know the terms but are unsure how tools, source material, and memory work together
  • Learners who want to understand the basics before developing agents

解决的问题

  • AI can write an answer without actually completing the work
  • You are unsure when to call a tool, consult source material, or save progress
  • Your design lists features but leaves checks and human approval points unclear

学完收获

  • Distinguish ordinary Q&A, ongoing assistants, fixed workflows, and agents using concrete scenarios
  • Choose the source material and tools a task needs, and define which actions are allowed
  • Plan task state, steps, and checkpoints, and decide whether to divide the work among multiple agents
  • Sketch a minimal agent design and use test cases to find gaps and risks

课程亮点

  • Customer follow-ups, refund decisions and meeting notes make the concepts concrete
  • Practice key decisions in each lesson by assessing scenarios and revising parts of a design
  • Ten lessons bring goals, source material, tools, state, and checks together in one design

全部章节

10 个章节

Lesson 1: What Is an AI Agent?

Distinguish ordinary Q&A, ongoing assistance, fixed workflows, and agents by how each one advances a task, and explain their action capabilities and human boundaries.

Lesson 2: Write Clear Task Instructions for an Agent

Review and locally rewrite agent instructions so they specify a checkable goal, an audience and responsibilities, source limits, action boundaries, and deliverables.

Lesson 3: Tools Let Agents Actually Do Things

Choose the fewest tools a task actually needs for its information and actions, and explain the call responsibilities, the permission checks, and the confirmation and failure-handling boundaries for high-risk actions.

Lesson 4: Why Agents Need Source Grounding

Identify the specific sources an answer depends on, choose how to obtain them, judge whether they apply, and stop unsupported conclusions when the evidence is missing or conflicting.

Lesson 5: Memory Helps Agents Carry Context Forward

Use what later tasks need in order to tell conversation history, current state, and lasting preferences apart, keep the information that is needed, and repair state that is stale, unverified, or applied to the wrong task.

Lesson 6: How Agents Complete Tasks Step by Step

Break a multi-step agent task into a minimal workflow whose steps depend on one another, with the key inputs, outputs, and state updates named, and place checkpoints before an error spreads or a risky action happens.

Lesson 7: When Multiple Agents Should Share the Work

Judge whether one agent is sufficient and, when the duties or the risks justify splitting the work, define the necessary roles, the routing, usable handoffs, and the final responsibility.

Lesson 8: Set Agent Permissions and Guardrails

Write guardrails with specific triggers and responses for the risks an agent faces, and judge from the rights, the consequences, and the validity of the approval whether to continue, to ask for confirmation, or to refuse.

Lesson 9: How to Tell Whether an Agent Is Reliable

Design representative fixed tests for an agent, locate the problems in observable process records, choose the smallest improvement that fits, and retest both the target issue and the capabilities that already worked.

Lesson 10: Draw Your First Agent Blueprint

Build a minimal agent blueprint for one scenario, check how its goal, sources, tools, state, workflow, boundaries, and evaluation connect, and repair the parts that do not link up.