Learn how AI agents observe their environment, choose actions, and complete tasks, starting with models, context, and tools. Eighteen lessons explain how source material, task state, memory, safety, and evaluation work together, then explore model training, real-time interaction, and collaboration among agents. Use this understanding to assess what an agent product can and cannot do.

适合人群

Beginners who have heard of AI agents and want to understand how their components work together
Product, operations, and technical professionals who need to assess whether an agent product is ready for use
Learners who want an overview before studying tool calls, safety, or collaboration among agents

解决的问题

It is hard to distinguish chat, fixed automation, and agents
You recognize the terms but cannot explain the respective roles of the model and the system around it

学完收获

Analyze an agent product in terms of its model, context, tools, and the system that runs it
Read a task trace and explain why the agent continues, stops, or hands control to a person
Explain how source material, task state, memory, and tools support the work in progress
Use permissions, actual results, and repeated tests to assess reliability and identify the cause of failures
Explain when training, real-time interaction, computer use, robot control, and multi-agent collaboration are useful

课程亮点

Eighteen lessons build from basic components to training, real-time interaction, and collaboration
Refund, research, web interaction, and robotics examples make abstract concepts concrete
Use failure traces and boundary cases to practice diagnosing causes and weighing design choices
Use a consistent set of questions to assess an unfamiliar product and decide which features it needs

全部章节

18 个章节

Lesson 1: What Is an AI Agent?

Analyze a specific AI product through its model, context, and tools, and distinguish a chat model, fixed automation, and an AI agent using observation, action, and feedback.

Lesson 2: How an Agent Completes a Task Step by Step

Read and complete an agent's observation–action trace, explaining the information gained, execution responsibility, and basis for continuing or exiting at each step.

Lesson 3: Choosing Between Workflows and Autonomous Agents

Choose a single call, fixed workflow, autonomous agent, or hybrid orchestration based on path predictability, environmental changes, and error costs, and define the model's control boundaries.

Lesson 4: How a Harness Manages an Agent

Locate model, context, tool, execution, verification, or state-management problems in an agent's failure trace and propose a verifiable Harness improvement.

Lesson 5: What Belongs in Context?

Select, organize, and label necessary context for a specific task, and distinguish prompt-expression, missing-information, information-organization, and model-capability problems.

Lesson 6: Manage Context and State in Long-Running Tasks

Choose status displays, caching, compression, isolation, and external state for long-running tasks, and design information-retention and safe-recovery rules.

Lesson 7: What Should an Agent Remember?

Decide whether information belongs in long-term memory and specify its subject, source, time, validity, update and deletion rules, and privacy boundaries.

Lesson 8: How a Knowledge Base Gives an Agent the Right Evidence

Inspect RAG through its materials, indexing, ranking, and answering layers, choose chunking and retrieval methods, and locate missing, misranked, or misused evidence.

Lesson 9: How Agents Choose and Call Tools

Design a tool contract serving both model decisions and system execution, and judge tool granularity, permissions, and the responsibilities of tools, Skills, and MCP.

Lesson 10: Controlling Risk Once an Agent Can Act

Design layered defenses along an attack or mistaken-action path, including content isolation, least privilege, argument checks, human confirmation, sandboxing, idempotency, verification, and rollback.

Lesson 11: Why Code Is a General-Purpose Agent Capability

Assign responsibilities to models, code, validators, and people according to task characteristics, and choose verification evidence matching the code-generated artifact.

Lesson 12: Proving That an Agent Actually Completed the Task

Design a repeatable agent evaluation covering environment, real completion evidence, process and safety criteria, repeated-run metrics, and failure diagnosis.

Lesson 13: How Post-Training Gives Models New Capabilities

Decide whether a capability problem calls for SFT, RL, distillation, or an external system improvement, and explain risks involving data, environments, feedback, generalization, and rollback.

Lesson 14: How Agents Become More Capable Through Experience

Extract, verify, and store real experience in a suitable form so it can be reused, evaluated, and reversed.

Lesson 15: From Waiting for Prompts to Responding in Real Time

Design event reception, deduplication, cancellation, interruption, state synchronization, and fast–slow processing for a real-time agent, and distinguish three kinds of real-time responsiveness.

Lesson 16: How Agents Operate Computers and Robots

Break interface and robot operations into observation, localization, action, and verification, and choose interfaces, visual methods, and hierarchical control based on environmental dynamics.

Lesson 17: When Multiple Agents Are Worth Using

Decide whether a task merits multiple agents and design context, collaboration topology, handoffs, conflict handling, and independent acceptance checks.

Lesson 18: A Framework for Understanding Any New Agent

Audit an unfamiliar agent through tasks, models, context, tools, Harness, evaluation, and learning locations, judging capability sources, necessary complexity, risks, and completion evidence.