AI terms explained, straight to the point.

An AI agent is a system built on a language model that decides on and carries out actions (such as querying tools, accessing files or calling APIs) to complete a task, instead of only answering questions.

Unlike a traditional chatbot, which only generates text in response to a question, an AI agent has access to tools and can plan multiple steps to reach a goal: look up information, fill in a form, send a message, run a script. Practical examples: a customer service agent that checks the CRM before replying, or a qualification agent that decides on its own whether a lead is ready for sales.

An agent squad is a group of specialized AI agents that collaborate, each responsible for one step of a larger process, to achieve a single business goal.

Instead of one generic agent trying to do everything, a squad splits the work: one agent researches, another writes, another reviews, another publishes, in an orchestrated way. It's the model used for more complex automations, such as a lead qualification squad (capture, enrichment, scoring and follow-up) or a marketing squad (briefing, creation, review and content scheduling).

RAG is a technique that retrieves relevant information from an external knowledge base before an AI model generates its answer, making answers more accurate and up to date than relying only on what the model learned in training.

A RAG system works in two steps: first it retrieves the documents or passages most relevant to the user's question, usually with vector search; then it uses those passages as context for the model to generate the final answer. It's the technique behind assistants that answer from a company's internal documentation, without retraining the model every time the content changes.

MCP is an open protocol created by Anthropic that standardizes how AI models connect to tools, data and external systems, so the same agent can access files, APIs and databases in a consistent way.

Before MCP, every integration between an AI model and an external tool required custom code. MCP defines a common interface: an MCP server exposes resources (data, tools, prompts) and any compatible client (such as Claude) can connect to it without a bespoke integration. In practice, this lets an AI agent access, for example, a database, Google Drive or a company's internal system using the same protocol.

LangChain is an open-source framework in Python and JavaScript for building applications with language models, offering ready-made components for RAG, agents, conversation memory and tool integration.

LangChain solves a recurring problem for anyone building with LLMs: connecting the model to data sources, managing conversation history, orchestrating multiple calls and integrating external tools. Instead of writing that logic from scratch, developers use LangChain components (chains, agents, retrievers, memory) to assemble applications faster: from chatbots that search documents to autonomous agents with multiple tools.

n8n is an open-source workflow automation tool with a visual interface that connects apps, APIs and AI agents without writing code, used to automate business processes end to end.

In n8n, a workflow is built by dragging and connecting nodes: a trigger (e.g. a message received on WhatsApp), processing steps (e.g. calling an AI agent to classify the message) and a final action (e.g. creating a card in the CRM). It's one of the most popular tools for giving an AI agent 'hands', connecting the model's reasoning to the real tools a company already uses.

Make is a visual cloud automation platform that connects apps and services through scenarios, similar to n8n but fully hosted and aimed at non-programmers.

Like n8n, Make automates processes by connecting triggers and actions across different tools (spreadsheets, CRMs, WhatsApp, email, AI agents). The main difference is that Make is a closed service hosted by the company itself, while n8n can be self-hosted, which often weighs on the choice between the two for business automation.

AI automation is the use of agents and language models to carry out steps of a business process that used to require human judgment, not just repetitive and predictable tasks as in traditional automation (RPA).

Traditional automation (RPA) follows fixed rules and breaks when exceptions show up. AI automation adds an agent that can interpret context, make decisions and handle variation: for example, classifying an ambiguous email, answering a customer question with nuance, or deciding the next step in a service conversation. That extends automation to processes that only a person could handle before.

Prompt engineering is the practice of deliberately structuring instructions, context and examples to get more accurate and consistent answers from an AI model.

A well-built prompt clearly defines the goal, the expected output format, the relevant context and, when needed, examples of the desired result. Structured prompt frameworks help turn loose prompts into repeatable, reliable instructions, reducing the variation in quality from one attempt to the next.

A webhook is a way for one system to notify another automatically and in real time when an event happens, sending the data over HTTP as soon as the event occurs, without anyone having to check manually.

Webhooks are the foundation of many automations and AI agent integrations: when a message arrives on WhatsApp, for example, a webhook instantly fires a workflow (in n8n, Make or a custom endpoint) that triggers the agent responsible for processing that message, without the system having to keep checking whether something new arrived.

An LLM is an AI model trained on large volumes of text to predict and generate natural language, the core technology behind assistants such as Claude, used both for conversation and for orchestrating agents.

Models such as Claude (Anthropic) are LLMs tuned for different levels of cost, speed and reasoning ability (for example, Haiku, Sonnet and Opus). Beyond answering questions, an LLM can serve as the 'brain' of an agent: interpreting instructions, deciding which tool to use and generating the next step of a task.

Agentic AI is the industry term for AI systems that plan and carry out tasks autonomously, with multiple steps and tool use, as opposed to models that only answer one question at a time.

According to Gartner, the share of enterprise applications with agentic AI is expected to jump from less than 1% in 2024 to 33% by 2028. In practice, that means systems that don't just suggest an answer but take the initiative to fetch data, call APIs and complete an entire workflow: like an agent that qualifies a lead, books a meeting and updates the CRM without human intervention at every step.

Reverse engineering is the process of analyzing finished software, without access to the source code, to understand how it works: which functions it has, how it stores data and which systems it talks to.

The work involves disassembling the compiled program into readable instructions, reconstructing pseudocode and recovering names, types and structures the compiler discarded. It's used in security, malware analysis, interoperability, software preservation and documentation of legacy systems. With projects such as REA, which went viral in 2026, AI agents started operating tools such as Ghidra and IDA, which sped up analysis but doesn't remove the need for human validation. Under Brazilian law, the Software Law protects the code, not the functionality, and legal risks appear when the analysis leads to copying code or bypassing protections.

A token is the unit of text a language model reads and writes: a piece of a word, a short word or a punctuation mark. It's also the unit used to bill AI model usage through APIs.

Before processing a text, the model splits it into tokens. In English, a token averages about three quarters of a word; other languages, such as Portuguese, usually need more tokens for the same content. Providers bill input tokens (what you send, including instructions and documents) separately from output tokens (what the model answers), so trimming instructions and avoiding resending entire documents lowers the cost of an agent in production.

The context window is the maximum amount of text, measured in tokens, that an AI model can consider at once: instructions, conversation history, documents sent and the answer itself.

Anything beyond the window's limit is left out: the model simply doesn't see it. The latest models accept hundreds of thousands of tokens, enough for entire books, but full windows cost more and can dilute the model's attention on what matters. That's why, instead of putting all of a company's documents into the context, enterprise assistants use RAG to fetch only the passages relevant to each question.

A hallucination is when an AI model generates false or made-up information that looks true, such as a figure, a quote or a reference that doesn't exist.

Language models generate the most likely text based on what they learned; they don't look up a fact database. When they don't have the information, they may fill the gap with something plausible. To reduce the risk in professional use: answer from company documents (RAG), require the source of every answer, instruct the assistant to say it doesn't know when the information isn't available and have people validate answers that drive important decisions.

Embeddings are numeric representations of the meaning of a text (or an image): lists of numbers in which similar content ends up close together, making it possible to search by meaning and not just by exact words.

An embedding model turns each piece of text into a vector. Questions and documents about the same subject produce similar vectors, even when they use different words. It's the basis of semantic search and RAG: the company's documents are converted into embeddings and stored in a vector database (such as Pinecone or Supabase's pgvector), and for each question the system fetches the closest passages to hand to the model.

Fine-tuning is the further training of an already trained AI model, using specific examples to change its behavior: the format of its answers, its tone of voice or its performance on a narrow task.

Fine-tuning teaches a model how to answer, not new facts that change often. For an assistant to know a company's documents and policies, the most common path is RAG, which fetches up-to-date information for each question without retraining anything. Fine-tuning makes sense when there are many high-quality examples of a repetitive task and well-written instructions aren't enough, and it requires care with the data used for training.

Prompt injection is an attack in which malicious instructions hidden in content (an email, a web page, a document) make an AI model ignore its original instructions and act on behalf of the attacker.

The risk grows with AI agents, which read external content and have access to tools: hidden text on a page could, for example, ask the agent to send data out. Prompt injection ranks first in the OWASP Top 10 for LLM applications. The main defenses are giving the agent only the permissions it needs, requiring human confirmation for sensitive actions, treating all external content as untrusted and monitoring what the agent does.

Generative AI is the category of artificial intelligence that creates new content, such as text, images, audio, video and code, from patterns learned in large volumes of data.

It differs from traditional AI, which classifies or predicts from data (for example, flagging whether a transaction is fraud). Assistants such as Claude, ChatGPT and Gemini are examples of generative AI built on language models (LLMs). In companies, the most common uses are customer service, marketing, document analysis and programming, and the next step is agentic AI, in which the model doesn't just generate content but also carries out tasks.