The model and how it answers
Language model (LLM): a program trained on a large amount of text that writes an answer word by word. It is the “brain” of the agent, but without instructions, knowledge and tools the model only talks. The other parts an agent is built from are shown on the seven-part diagram.
Token: a piece of text the model counts in: a word, part of a word or a symbol. Cloud models charge for tokens in and out, so long instructions and documents cost money with every answer; monthly costs are worked out on the page on the cost of running an agent.
Context window: how much text the model sees at one time, its working memory. The instruction, the dialog history and the documents found all go here, and what does not fit, the model ignores.
Temperature: a setting for the randomness of an answer. A low value gives restrained, similar answers and a high one gives varied answers; even at zero, answers to the same question may differ, so an agent is tested on a set of questions.
Instructions and knowledge
System prompt: a permanent instruction to the agent: who it is, what it does, what it does not do and how it answers. You can edit it without a programmer, and it affects the answers most; a template is on the page on the system prompt.
RAG (search in documents before answering): the agent first finds fragments in your document base, then answers from them. This is how it works with a price list and policies without training the model; see the step-by-step build in the RAG guide.
Vector database: a store where document fragments are kept as vectors, sets of numbers that reflect meaning. With them the agent finds text close to the question in meaning, even if the words are different.
Fine-tuning: extra training of a ready model on your examples, so that it takes on a style or a format. A business agent rarely needs it: RAG gives knowledge, and the system prompt sets behavior.
Actions and connections
Tool (function calling): an action the model can request: find a customer in the CRM, create a request, add an entry to a calendar. The model decides when to call a tool, and your program makes the call and returns the result. How to give the agent tools without extra rights is covered separately.
MCP (Model Context Protocol): an open standard for connecting AI to files, databases and services, often compared to a USB-C port. Anthropic opened the protocol on 25 November 2024, and on 9 December 2025 it was handed to the Agentic AI Foundation at the Linux Foundation; step-by-step setup is in the MCP server guide.
Agent loop (ReAct): the model alternates reasoning and action: it decides which step is needed, calls a tool, looks at the result and decides again. Most agents work on this principle, described in 2022; a diagram of the loop is on the page on how an AI agent works.
Agent, workflow and team of agents
Workflow (automation scenario): the steps are set in advance in code or in a builder, and the model carries out single steps but does not choose the order. For repeatable tasks it is often enough, and it breaks less often.
AI agent: a system where the model chooses the steps and tools by itself until the task is solved. Anthropic advises starting with the simplest solution and moving to an agent only when a workflow is not enough.
Multi-agent system: several agents with different roles that pass tasks to each other: one collects data, another writes the answer, a third checks. It is needed when a process runs through several departments.
Agent memory: what the agent remembers between conversations: a customer’s name, past orders, agreements. Short-term memory lives in the context window and long-term memory in a database, while company rules set how long personal data is kept.
Where the model runs
Open model (open weights): a model you can download and run yourself. For example, DeepSeek R1 came out on 20 January 2025 under the MIT license, and the Qwen 3 family came out on 29 April 2025 under the Apache 2.0 license; the licenses of open tools are listed on the page on free agents.
Local model: a model that runs on your own computer or server, not in someone else’s cloud. It is most often run through Ollama, a program under the MIT license, and customer data does not leave the company.
Agent SDK: a set of libraries a programmer uses to build an agent in code: the loop, the tools and the hand-off of a task between agents. An example is the Claude Agent SDK, the new name Anthropic gave the Claude Code SDK on 29 September 2025.
Testing, security and launch
Hallucination: a confident answer that is not in your data, such as an invented price or term. RAG, a ban on answering without a source and testing on a set of questions reduce it.
Test set: real customer questions with correct answers, on which the agent is checked before launch and after every edit. Without it you cannot say whether the agent became more accurate.
Prompt injection: someone else’s instruction hidden in an email, a document or a website that the agent may take for a command. The defense: minimal rights, human confirmation of risky actions and a log.
Human in the loop: a rule that the agent does not take an important action (a payment, a refund, sending a contract) without an employee’s confirmation. If the agent sees customer data, 152-FZ, Russia’s personal data law, adds its own control: data of Russian citizens is collected and stored in databases in Russia, as described on the page on data security.
Pilot and metric: a launch of the agent on one process, with a “before and after” number in the contract. The metric is reviewed every week, and it decides whether to expand the agent or stop it.