An agentic system gives a model a bounded way to gather information, choose an action, and observe what happened. The model supplies judgment where fixed rules are too rigid. Ordinary code still owns permissions, validation, budgets, durable state, and irreversible effects.
- AI, ML, and deep learning
Machine learning is within artificial intelligence; deep learning is within machine learning.
- Generative models
Generative AI produces content such as text, images, audio, and code. LLMs are a family of models for token sequences.
- Agent system
Combine a model with an execution loop, tools, state, policy, and stop conditions. The system’s behavior is larger than the model.
An agentic system includes an application around a model; it is not another name for an LLM.
How to read this model
A nested storage cabinet beside an independent agent workbench. The object is a teaching analogy. Inspecting a part highlights its role; it does not execute or simulate the software. Cabinet trays are records, workpieces are data, and latches are checks. Unless an explanation states a quantity, compartments and sample workpieces do not denote data counts. Part dimensions are not a quantitative scale.
That division of work is the central idea in this handbook. A model may suggest that an incident needs a Kubernetes restart. It must not gain cluster-admin access merely because it produced a convincing explanation.
- Assemble context
The user goal and constraints guide application-selected context.
- Model proposal
The model can propose restartDeployment with name orders and namespace prod. It does not grant itself production authority.
- Validate and authorize
Application code decides whether the proposed operation is valid and permitted.
- Capabilities and observation
Tools read or change state; retrieval supplies evidence; a subagent runs bounded work. Record observations for the next turn.
Structured does not mean authorized or safe.
How to read this model
A proposal press inside a guarded workbench with separate capability bays. The object is a teaching analogy. Inspecting a part highlights its role; it does not execute or simulate the software. Cabinet trays are records, workpieces are data, and latches are checks. Unless an explanation states a quantity, compartments and sample workpieces do not denote data counts. Part dimensions are not a quantitative scale.
Follow the dependency chain
- Language models
Understand tokens, context, attention, generation, embeddings, and structured output.
- Retrieval
Give a model current evidence without pretending that retrieval makes every answer true.
- Agents and tools
Place model decisions inside a loop with explicit actions, observations, and stop conditions.
- Context and memory
Choose what the model sees now and what the application keeps for later.
- Subagents
Delegate only when isolation, specialization, or parallel work earns the extra coordination cost.
- MCP
Connect hosts to tools and context through a versioned protocol with explicit trust boundaries.
- Production engineering
Design for unsafe input, partial failure, traces, evaluation, latency, and cost.
Keep five boundaries visible
Most production failures become easier to explain when you ask which boundary was blurred.
- The model proposes. A model emits text or structured data. It does not grant itself authority.
- The application decides. Deterministic code validates the proposal, checks policy, and chooses whether execution may continue.
- A tool acts. A named operation touches a database, API, file, browser, queue, or other environment.
- State records what happened. Durable state, not conversational prose, records commitments and recovery points.
- Evaluation checks the whole path. A fluent final answer can hide a wrong tool, an unsafe argument, or a duplicate side effect.
An agent framework may package these boundaries, but it cannot remove them.
Choose the right amount of autonomy
Use ordinary code when the inputs and rules are known. Use a workflow when the path is known but some steps need model judgment. Use an agent when the next useful step depends on evidence discovered during execution.
More autonomy creates more possible execution paths. Each path needs permissions, limits, observability, and tests. The most capable design is often the one that gives the model the smallest decision that still solves the problem.
How these pages are written
Concept pages explain one mechanism and its trade-offs. Tutorials build a working example. Reference pages give exact terms, comparison tables, and checklists. Design walkthroughs apply the material to a complete system.
Start with the learning path that matches your background. Use the question coverage index when you want to trace the original 560-question study list into the handbook.
Sources
- Attention Is All You Need
- ReAct: Synergizing Reasoning and Acting in Language Models
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
- Model Context Protocol specification
- NIST Artificial Intelligence Risk Management Framework
Last verified: September 23, 2026.