Use the terms in English to recognize them in official documentation and materials. The definitions are operational: they seek to help you choose and explain, and not memorize words.
Agents and reasoning
| Term | What does that mean? | Don't confuse him with |
|---|---|---|
| Agent | A system that decides actions to an goal, using a model, tools and a cycle. | Any regular chatbot or pipeline. |
| LLM | A model of language that produces an output from a context. | A function performer with automatic access to your system. |
| Inference | Performance of model to produce an output. | Training model. |
| Tool | Incusable capacity with an entry / exit contract. | A written instruction that itself grants permissions. |
| Tool calling | Structured output that requires using a tool. | The effective execution of that tool. |
| Agent loop | A cycle that coordinates decisions, tools, observations and stop. | A single call model. |
| ReAct | A pattern that modifies reasoning and actions and incorporates observations. | an obligation to show internal reasoning to the user. |
| Chain | A problem's decomposition at a step of reasoning. | A tool for viewing external data. |
| State | Information that the system keeps during the flow. | Change model weights. |
| Context window | Quantity of context that a call to the model can process. | Unlimited permanent memory. |
Orchestra and integration
| Term | What does that mean? | Useful distinction |
|---|---|---|
| LangChain | Framework with interfaces and components for model applications. | The application retains responsibility for permissions and validation. |
| LCEL | How to compose components into LangChain chains. | The composition operator does not make a string autonomous. |
| MCP host | Application that contains and coordinates MCP customers. | It doesn't have to be the model. |
| MCP client | Component that keeps a connection with a MCP server. | Host and client have different responsibilities. |
| MCP server | It exposes skills to customers as per protocol. | It doesn't replace the agent's cycle by itself. |
| Resource | Content or data expressed by a MCP server. | That's an executable tool. |
| Propt | Interplay plate offered by a MCP server. | Access to all system instructions. |
| Responses API | API OpenAI that integrates responses, tools and context options / state. | Do not run an automatic local Python function. |
| Agents SDK | The orchestration library of agents, tools, handoffs and other capabilities. | A rental service from your application. |
| Handoff | Transfer of conversational monitoring to a specialist agent. | Use another agent as a tool and get back to the main agent. |
Security and operation
| Term | What does that mean? | Question that should be made |
|---|---|---|
| Guarrail | Control that valids or restricts some of the flow. | At what point and about what contents are run? |
| Propt injection | Malicious instructions included in unreliable inputs or contents. | Am I processing outside data as authorized instructions? |
| Last privilege | Grant only necessary permissions. | Can this tool do more than the task requires? |
| Human--in-the-loop | Human intervention in concrete decisions. | Which sensitive operation requires confirmation? |
| Tracing | Record of operational events of the flow. | Can I rebuild calls, mistakes and times? |
| Runtime | The environment that runs and coordinates the agent's conduct. | Who keeps sessions, runs tools and applying boundaries? |
| Evaluation | A systematic check with established cases and criteria. | Am I measuring accuracy, safety and cost as well as fluidity? |
Data and Recovery
| Term | What does that mean? | Useful distinction |
|---|---|---|
| Embedding | Numeric representation of contents. | It's not an original document or a guarantee of truthfulness. |
| Chunk | Fragment of a document used in recovery. | An arbitrary cut may lose necessary context. |
| Vector search | Search by similarity between vector performances. | Symility doesn't mean accuracy factual. |
| RAG | Reclaiming context and using it to base a generation. | It doesn't amount to training or adjusting the model. |
| Select AI | Capacities to use natural language with AI services / models and Oracle data. | The generation of SQL does not give new permissions. |
| Select AI Agent | A framework for defining agents, tasks and tools close to data. | It does not mean that the LLM weights are run within the base. |
| Private Agent Factory | Environment to build agents and flows with visual interface. | The visual configuration does not remove the need for governance. |
A Test of Understanding
Explains the difference between tool calling, MCP y run with an order from a tent. A solid response identifies structured request, an integration protocol and an environment that implements and coordinates work. If you use all three terms as synonyms, come back to MCP module.