Understand the concepts
Start with prerequisite concepts, causal explanations and counterexamples, and be able to explain them clearly in your own words.
Start with model call →Knowledge catalogChoose core direction and segmented content
SYSTEMATIC LEARNING
Agent application development: knowledge map
Starting from business tasks, the system learns the interface calling, tool integration, retrieval, memory, reliable execution, evaluation and launch of large model applications. Go in-depth step by step along implementation, faults, and engineering trade-offs, and then use interview questions and experiments to check your understanding.
Take each concept deeper into engineering practice
Developers who are new to Agent should first learn the basics, and continue to study faults and trade-offs after gaining implementation experience. Choose a starting point based on your familiarity with this knowledge point.
Build the implementation
Read the implementation and code, complete hands-on tasks, and check the acceptance requirements one by one.
Go to Execution and Orchestration →Diagnose failures
Use follow-up questions and transfer scenarios to analyze crashes, retries, and changing conditions.
Learn to perform reliably →Explain trade-offs
Compare options within business constraints and explain costs, risks, and applicability boundaries.
Enter system design →There are 8 four-level courses with separate lessons and tasks. Units marked as core lessons provide concepts, follow-up questions, and transfer cases; they have not yet been expanded into four-level courses.
Core area 01
Model calls and context
Based on application needs, understand interface returns, context capacity and model capability boundaries. Pre-training and fine-tuning only talk about how they affect model selection, knowledge updating and evidence use; the tasks in this direction revolve around request assembly, response processing and application solution selection.
Questions that run through this direction:Where does the model obtain information, what stage is the output at, and what judgments can the program make based on this?
After learning, what should you be able to explain?
- Calculate input, output and context margins and explain truncation reasons
- Select retrieval, tools or fine-tuning solutions based on knowledge update, evidence and behavioral needs
- Identify different processing paths for text, tool requests, rejections, and incomplete responses
- Explain the relationship between valid structure, correct facts and business completion
- Your first model call and response contractMove from ordinary HTTP APIs to model APIs: build requests, distinguish response types, correlate tool results, and verify answers against execution evidence.
Four-level course · Separate lessons and tasks for foundation, implementation, debugging, and design
Practice · 2 linked exercises · Start answering → - Prompts, structured output, and iterative acceptance checksUse a standard return, a product exception, and a missing policy to connect task instructions, output constraints, evidence citations, and version comparisons.
Four-level course · Separate lessons and tasks for foundation, implementation, debugging, and design
Practice · 3 linked exercises · Start answering → - Context and generation budgetsCalculate the capacity of a complete request and connect input size, output budget, growth in the next turn, and evidence retention.
Four-level course · Separate lessons and tasks for foundation, implementation, debugging, and design
Practice · 3 linked exercises · Start answering →
Core area 02
Task contracts and execution orchestration
Autonomous planning must fall within controllable task boundaries. Task contracts, status, budgets, and transitions determine how the system moves forward, when it ends, and how multiple branches work together.
Questions that run through this direction:How to organize the model's next-step recommendations into a constrained and acceptable execution process?
It is recommended to understand first:Model calls and context
After learning, what should you be able to explain?
- Define execution boundaries with task inputs, constraints and acceptance
- Compare when state graphs, autonomous loops, and multiple agents are appropriate
- Explain invariants in concurrent merging, replanning, and context compression
- Control and completion criteria in agent loopsStarting from the tool invocation loop, design termination conditions, budgets, cancellations, and verifiable completion statuses.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Combine deterministic workflows with bounded explorationChoose a recoverable orchestration structure around defined processes, dynamic exploration, checkpoints, and manual approvals.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Algebra and business semantics of parallel state mergingExamine state modeling, reducers, concurrent merging and message deduplication.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Task contracts and independent acceptance checksMake the execution environment, permissions, status, evidence and completion judgment of the model into a project contract.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Benefits and coordination costs of multiple agentsExamining task parallelizability, coordination costs, evidence merging, and controlled experiments.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Separate plan revisions from external factsExamine dynamic plans, completed side effects, dependencies, and acceptance invariants.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Transfer responsibility and permissions during task handoffExamining Handoff, Identity Propagation, Minimal Context, and Responsibility Boundaries.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering →
Core area 03
Tool contracts and external actions
Tools allow models to influence real systems. Parameters, identities, protocols, versions, and business idempotency jointly determine whether a tool request can be safely understood, executed, and verified.
Questions that run through this direction:How does a seemingly legitimate tool request become a verifiable business action?
It is recommended to understand first:Task contracts and execution orchestration
After learning, what should you be able to explain?
- Distinguish between structure verification, business verification and authorization
- Explain why timeouts, disconnections and retries require result verification
- Design tool collections, compatible versions, and delivery methods for big results
- Tool calls: structure, authorization, and business contractsStart with a well-formed request that crosses a tenant boundary, then examine each check needed to turn a tool request into an executable action.
Four-level course · Separate lessons and tasks for foundation, implementation, debugging, and design
Practice · 3 linked exercises · Start answering → - Separate discovery from execution in large tool catalogsExamine tool catalog, dynamic discovery, recall quality, and server-side authorization.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Delivery conditions and shared deadlines for parallel resultsExamine asynchronous concurrency, deadlines, cancellation propagation, and partial results.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Computation, references, and consistent snapshots for large resultsExamine output contracts, paging, artifact references, and deterministic calculations.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Reconnect transport, recover messages, and reconcile business outcomesExamine the differences between stdio, Streamable HTTP, session recovery and business idempotency.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering →
Core area 04
Retrieval, evidence, and long-term memory
Answer quality depends directly on whether the model receives relevant, trustworthy information it is currently authorized to use. Retrieval and memory share source, scope, version, and revocation boundaries, while having different read, write, and evaluation mechanisms.
Questions that run through this direction:How do you put correct and acceptable information into context and know when you can't answer?
It is recommended to understand first:Model calls and context, Tool contracts and external actions
After learning, what should you be able to explain?
- Positioning errors along recall, sorting, assembly and generation
- Handle chunking, query rewriting, reference and evidence conflicts
- Managing memory writing, retrieval, scoping, and undoing
RAG and retrieval engineering
- From document ingestion to cited RAG answersRun Markdown parsing, replay recorded embedding vectors, apply SQL permission filters, and inspect BM25, vector retrieval, RRF, and answer acceptance checks.
Four-level course · Separate lessons and tasks for foundation, implementation, debugging, and design
Practice · 5 linked exercises · Start answering → - RAG evidence flow and failure diagnosisFollow the same evidence through retrieval, ranking, context assembly, and the answer to see where the information needed for a correct answer can be lost.
Four-level course · Separate lessons and tasks for foundation, implementation, debugging, and design
Practice · 4 linked exercises · Start answering → - Consistent authorization across RAG and derived dataEnforce authorization throughout retrieval, reranking, context assembly, caching, and source access.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Definitions, granularity, and execution boundaries in natural-language queriesExamine the semantic layer, controlled SQL, read-only execution, and result review.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Entity disambiguation and constraint preservation in query rewritingExamining multiple rounds of reference, entity linking, clarification, and original constraint preservation.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering →
Memory management
- Memory scopes and trusted authorization contextDesign memory scope, authenticity, permissions and recall order as data contracts, and vector similarity is only responsible for sorting.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Admission, provenance, and fact correction for long-term memoryExamine memory access, sources, validity periods, and error correction.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Objects, sources, and bitemporal reasoning in memory conflictsExamine fact versions, validity times, source priority, and conflict presentation.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Hard filters, value ranking, and causal evaluation in memory retrievalExamining memory retrieval stratification, relevance, authority, temporality, and counterfactual assessment.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Revoke memories, clean up derivatives, and prevent stale data from returningConnect deletion to recovery and handle old state with undo versioning, provenance verification, physical cleanup, and failure testing.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering →
Core area 05
Reliable execution and task recovery
Long tasks will encounter process interruptions, duplicate messages, approval waits, and status upgrades. Only by understanding the fault window and status boundaries can the recovery process be checked with external business effects.
Questions that run through this direction:How to continue without losing business facts after a task is interrupted, replayed or taken over?
It is recommended to understand first:Task contracts and execution orchestration, Tool contracts and external actions
After learning, what should you be able to explain?
- Distinguish between checkpoint, idempotent, lease and business completion status
- Handle unknown results, duplicate approvals, and compensation failures
- Design state migration, message delivery and multi-tenant queuing strategies
- Idempotency, unknown outcomes, and task recoveryUse separate caller and service databases to simulate a successful commit with a lost response, and examine what checkpoints can guarantee.
Four-level course · Separate lessons and tasks for foundation, implementation, debugging, and design
Practice · 5 linked exercises · Start answering → - Atomic intent and duplicate delivery with the outbox patternExamine double-write issues, transaction outboxes, duplicate deliveries, and recovery scans.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Structural and business compatibility during workflow upgradesCheck workflow version, persistence state schema and playback compatibility.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Committed effects and conditional compensation in sagasExamine cross-service transactions, compensation conditions, irreversible actions, and human intervention.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Fair scheduling, resource quotas, and backpressure across tenantsExamine multi-tenant queuing, fair scheduling, backpressure and multi-dimensional quotas.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering →
Core area 06
Evaluation, acceptance, and iteration
A successful demonstration is not enough to prove that the system is reliable. Results, processes, test sets, repeatability, and human judgment need to be designed separately to detect improvements and degradations caused by changes.
Questions that run through this direction:How to use interpretable evidence to determine whether the Agent is doing the right thing, is doing it steadily, and is worthy of being launched online?
It is recommended to understand first:Task contracts and execution orchestration, Tool contracts and external actions
After learning, what should you be able to explain?
- Establish acceptance conditions for results and processes
- Identify data leakage, grader bias, and inconsistent measurement definitions
- Compare different versions using fault testing and regression
- Agent evaluation: outcomes, constraints, and evidenceStart with a well-written report that accessed unauthorized data, then design an acceptance contract whose results can be verified and failures diagnosed.
Four-level course · Separate lessons and tasks for foundation, implementation, debugging, and design
Practice · 5 linked exercises · Start answering → - Failure windows and recovery invariantsExamine failure points, business invariants, deterministic tool stand-ins, and recovery acceptance.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Release and rollback of behavioral configurationsExamine behavioral versions, regression thresholds, shadow traffic and online attribution.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering →
Core area 07
Security boundaries and production operations
After launch, maintain permissions, quality, cost, and runtime state together. Security controls and observability belong at execution boundaries; caches, fallback models, and streaming connections must preserve the same business contract.
Questions that run through this direction:How to keep the system controllable in the presence of real users, failures, and untrusted input?
It is recommended to understand first:Tool contracts and external actions, Reliable execution and task recovery, Evaluation, acceptance, and iteration
After learning, what should you be able to explain?
- Implement approval, sandboxing and minimum permissions into action execution
- Locating performance, cost and quality issues along operational evidence
- Maintain consistency in caching, downgrading, outage recovery, and incident handling
- Task causality and resource attributionString business tasks, model requests and tool calls into queryable links, and use quality regression to supplement operating indicators.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Approval targets and execution-time authorizationImplement permission checks at the tool execution boundary, bind approval to immutable parameters and resource versions, and reject prompt injection to expand authorization.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - External data and execution-permission boundariesExamine indirect prompt injection, data and command separation, tool permissions, and outbound control.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Isolate resources and capabilities when executing codeExamine files, networks, processes, credentials, and resource limits.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Capability and status contracts for fallback modelsExamine multi-model adaptation, capability matrices, failure classification, and degradation.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Semantic and authorization equivalence for cache reuseExamine different caching tiers, permission versions, latency and real savings.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Reconnecting clients and durable task stateExamine task and connection decoupling, event sequence number, reconnection and clear final state.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Incident containment, evidence, and causal diagnosisExamine stop loss, version attribution, tiered metrics, privacy, and review.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering →
Core area 08
System design and engineering trade-offs
Comprehensive scenarios put the previous mechanisms under common constraints. Understanding the business goals, acceptable risks and acceptance methods enables you to select the appropriate architecture and explain under what conditions it needs to be adjusted.
Questions that run through this direction:How do you combine these mechanisms into deliverable systems based on mission, risk, and team conditions?
It is recommended to understand first:Retrieval, evidence, and long-term memory, Reliable execution and task recovery, Evaluation, acceptance, and iteration, Security boundaries and production operations
After learning, what should you be able to explain?
- Derive architecture from enterprise tasks rather than stacking components
- Explain the trade-offs between framework selection, customer service actions, R&D tasks, and research release
- Explain the project experience with the design and evidence that I am responsible for
- A controlled task chain from evidence to publicationConnect identity, knowledge permissions, structured tools, persistence tasks, and audit releases into a recoverable system.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Match orchestration abstractions to business responsibilitiesUse state complexity, recovery requirements, tool contracts, and team constraints to make trade-offs, rather than choosing based on framework popularity.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Bind code changes to acceptance evidenceExamine code base understanding, isolation execution, acceptance, regression and manual review.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Separate natural-language requests from transaction stateExamine intent understanding, business rules, authorization and traceable execution.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Bind claims to sources and verification scopeExamine research plans, source verification, evidence ledgers, experimentation and publication thresholds.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Tenant isolation and continuous authorization during executionExamine the control plane, execution plane, data isolation, capacity, and progressive delivery.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering → - Evidence and conditional reasoning about project capabilitiesExamine the project review, personal contribution, quantitative evidence and technical choices.
Core lesson · Concepts, follow-up questions, and transfer cases
Practice · 1 linked exercises · Start answering →
How can I understand it and draw inferences?
- First use core principles to establish causal relationships, and then read about implementation and trade-offs.
- Follow the continuous question and answer process to see how the plan changes after the constraints change.
- Compare transfer cases to identify shared invariants and different assumptions.
- When you need to check your understanding, do short answers or small experiments and record your explanations.