A beginner-friendly course for building a clear mental model of AI agents. Starting with models, context, and tools, it explains agent loops, harnesses, memory, RAG, safety, evaluation, post-training, real-time interaction, and multi-agent collaboration.
Analyze real AI products in terms of model, context, and tools, then use observation, action, and feedback to distinguish chat models, fixed automation, and AI agents.
Read and complete an agent's observation-action trace, explaining the new information, execution responsibility, and continue-or-exit basis at each step.
Choose a single model call, fixed workflow, autonomous agent, or hybrid orchestration based on path predictability, environmental change, and the cost of errors, while defining what the model may control.
Use failed agent traces to locate problems in the model, context, tools, execution, verification, or state management, then propose one testable harness improvement.
Select, organize, and label the context needed for a task, while distinguishing prompt wording, missing information, poor information organization, and model capability problems.
Choose status summaries, caching, compression, isolation, and externalized state for long-running tasks, and design rules for retaining information and recovering safely.
Decide whether information belongs in long-term memory, and define its subject, source, time, validity, update and deletion rules, and privacy boundaries.
Inspect RAG across source material, indexing, ranking, and answer generation; choose chunking and retrieval methods; and locate missing, misranked, or misused evidence.
Design a tool contract that supports both model decisions and system execution, and judge tool granularity, permissions, and the boundaries among tools, Skills, and MCP.
Trace attacks and mistakes to design layered defenses using content isolation, least privilege, parameter checks, human confirmation, sandboxes, idempotency, verification, and rollback.
Assign responsibilities among the model, code, verifiers, and people based on the task, then choose verification evidence that fits the output produced by code execution.
Design repeatable agent evaluations that cover the environment, real completion evidence, process and safety criteria, repeated-run metrics, and failure diagnosis.
Decide whether a capability problem belongs in SFT, RL, distillation, or an external system, and explain the associated data, environment, feedback, generalization, and rollback risks.
Extract, verify, and store real experience in the right form so it remains reusable, evaluable, and reversible.
Design event intake, deduplication, cancellation, interruption, state synchronization, and fast-slow processing for real-time agents, and distinguish three forms of real-time speed.
Break interface and robot operation into observation, localization, action, and verification, then choose APIs, visual methods, and layered control based on how dynamic the environment is.
Decide whether a task merits multiple agents, then design context boundaries, collaboration topology, handoffs, conflict handling, and independent acceptance checks.
Audit an unfamiliar agent through its task, model, context, tools, harness, evaluation, and learning location, then judge the source of its capabilities, necessary complexity, risks, and completion evidence.
