Overview
Course facts: 6 lessons, 50 minutes, no sign-up, no payment. The course is for an employee or a manager.
Lesson 1: what is an AI agent, and how is it different from a chatbot?
A chatbotchatbot. A chatbot is a program that exchanges messages with a person in a chat window: on a website, in a messenger or in an app. answers a question and waits for the next one. An AI agent receives a task and carries it out step by step: it looks up data, calls the needed programs, checks the result and decides what to do next. The difference is not in the model. It is that the agent is allowed to act.
By the OpenAI guide “A practical guide to building agents” (April 2025), an agent has three parts:
- Model. The neural network that reasons and makes decisions.
- Tools. The functions and programs through which the agent acts: search in a database, a write to a CRM, sending a letter.
- Instructions. Explicit rules and limits: what the agent does, what it does not do and when it calls a person.
An agent works in a loop: gather context, take an action, check the result, repeat. This approach was described in 2022 under the name ReActReAct. A ReAct agent is an AI agent that works in turns: it reasons, calls a tool, looks at the result and reasons again.: the model alternates reasoning and an action with a tool. Most AI agents work on it.
Not every system with AI works as an agent. In the article “Building effective agents” (December 2024), Anthropic splits such systems into two kinds:
- Workflow. The steps are set in advance by code: receive a letter, extract details, write to a table. The model does single steps but does not choose their order.
- Agent. The model itself chooses which steps to take and which tools to call until the task is solved.
An example: a workflowAI workflow vs agent. An AI workflow and an AI agent are two ways to build an AI system: in a workflow, code sets the steps in advance, and an agent chooses the next step and its tools while it works. answers a request from the site by a template and passes it to a manager. An agent reads the request, finds the client in the CRM, looks at the order history, asks for what is missing and creates a task for the manager with a ready offer.
Lesson 2: which tasks should you give an agent, and which not?
Start with a simple solution and add complexity only when needed. This is the main advice of Anthropic. Sometimes it means you do not need an agent at all. If the steps of a task are always the same, a workflow or ordinary automation is more reliable and cheaper.
An agent is justified when the task has at least one of three features (by the OpenAI guide):
- Complex decisions and exceptions. For example, deciding whether to approve a refund to a customer.
- Heavy rules that are costly and slow to keep in code. For example, checking a supplier against dozens of conditions.
- A lot of unstructured data: letters, scans, correspondence, contracts.
Do not give an agent:
- Tasks with identical steps: moving data from a form to a table, a scheduled mailing.
- Decisions that nobody is ready to answer for: payments, refusing a client, dismissing a person.
- Tasks without data: an agent will not learn prices and rules that are written nowhere.
Why this matters: by a Gartner forecast (June 2025), more than 40% of agentic AIagentic AI. Agentic AI is an approach where a neural network receives a goal and carries the work to a result on its own. It plans the steps, calls software tools and checks what came out. projects will be closed by the end of 2027 because of cost, unclear benefit and weak risk control. All three causes are set at the start, when you choose the task.
Check a task with three questions:
- Can you describe the result in one sentence and check it?
- Are the data and access you need available?
- What happens if the agent makes an error, and who will notice?
Lesson 3: what is an agent built from?
An agent is built from a model, tools, knowledge and memory.
- Model. Build the prototype on a strong model, to see what quality is possible at all. Then try smaller models where the result stays acceptable. That makes work cheaper and faster.
- Tools are of three kinds: for data (search in a knowledge base, reading a CRM, tables and mail), for actions (create a deal, send a letter, book a client) and for control (pass part of the task to another agent).
- MCPMCP. The Model Context Protocol (MCP) is an open standard that lets AI applications connect to tools and data: files, a CRM, databases and online services.. To avoid writing a connection to every service again, Anthropic opened the MCP protocol (Model Context Protocol) in November 2024. It is a single way to connect AI to data and services. In December 2025 the protocol was passed to the Agentic AI Foundation at the Linux Foundation, so no single company owns it.
- Knowledge. A model does not know your prices, contracts and rules. For an agent to answer from them, use RAGRAG. Retrieval-augmented generation (RAG) is a scheme in which an AI first finds the right passages in your documents and then answers using them.: before it answers, the agent searches your document base for the needed pieces and answers from them. You do not need to retrain the model. If the base is small, under 500 pages, you can put it all into the request without search.
- Memory is of two kinds: short-term (the history of one conversation) and long-term (data about a client or a task that is kept between conversations). The full history may not fit into a request, and most models work worse on long text. So outdated items are removed from memory.
- Instructions are written like an assignment to a new employee: take the existing rules, split the task into steps, set a concrete action for each step and describe unusual cases.
Lesson 4: what do you build an agent on?
For first agents you do not need to program. Open builders let you assemble an agent from blocks in a browser and install it on your own server:
- n8n. An automation platform with AI agents. Its license is fair-code: you can install it on your server and change it, but you cannot resell it as a cloud service.
- Flowise. A visual agent builder under the Apache 2.0 license.
- Dify. A platform for AI applications and agents under Apache 2.0 with two conditions: you may not run a multi-tenant cloud service like Dify without permission, and you may not remove the logo from the console.
When you need code, take open libraries: LangChain and LangGraph for Python and JavaScript, and CrewAI for teams of several agents. All three are under the MIT license.
For the model, open modelsopen weights model. An open weights model is a neural network whose parameter files you can download and run on your own hardware. that you install on your own computer or server suit work tasks well. Examples are DeepSeek R1 (MIT license) and the Qwen 3 family (Apache 2.0 license). Ollama is a convenient way to run them. An open model on your own server also solves the data question: correspondence, documents and customer details do not leave the company.
How to choose:
- One agent, clear steps, ready connections to services: n8n.
- A chat with answers from your documents: Dify or Flowise.
- Complex logic, several agents, your own product: libraries and a developer.
Lesson 5: how do you build a first agent in one evening?
Take one task that you do by hand every week. For example, sorting incoming letters: understand what a letter is about, find the client and prepare a draft reply.
A build plan on n8n and an open model:
- Write the task in one sentence and describe what a good result looks like.
- Install Ollama and download a model that can call tools. For example, qwen3:8b weighs 5.2 GB, and if memory is short there is qwen3:4b at 2.5 GB.
- Start n8n with the command npx n8n and open it in the browser.
- Add the AI Agent node and connect the model to it. The agent needs at least one tool, or the node does not work.
- Write the instruction: a role, the steps, the limits and what to do if data is missing.
- Connect a data tool first, such as a search in a client table, and only then an action tool.
- Test the agent on 20–30 real examples and write down where it made mistakes.
Two traps of a first agent:
- Temporary storage. The Simple Vector Storevector database. A vector database is a store that keeps texts, images and products as embeddings and quickly finds records close in meaning to a query. node in n8n keeps data only in memory, so it is lost after a restart. It is fine for tests. For real work you need a real database.
- Extra tools. Each extra tool gives the agent one more chance to err. Add them one at a time and check after each.
Lesson 6: how do you control an agent and keep it safe?
An agent acts by itself, so its errors cost more than a chatbot’s. OWASP, an international application security community, calls this risk excessive agency. Its causes are extra functions, extra rights and extra autonomy.
Five rules of control:
- Only the needed tools. An agent that answers customers does not need access to accounting.
- Minimal rights. If an agent only reads the CRM, do not give it the right to delete deals.
- A person confirms important actions. Payments, large refunds and order cancellations pass a human check until the agent proves reliable.
- Handover to a person. If the agent fails after several tries, the task goes to an employee.
- An action log. Every step of the agent is recorded, so you can study an error later. Open tools exist for this, for example Langfuse.
The company is responsible. In February 2024 a tribunal in British Columbia, Canada, found Air Canada responsible for a wrong answer of a chatbot on its site. The airline’s argument that the chatbot answers for itself was rejected.
Data: if the agent works with customer data, check where the model runs and where the knowledge baseAI knowledge base. An AI knowledge base is a set of selected and prepared company documents that an AI assistant searches for answers: rules, instructions, a product catalog and answers to common questions. is stored.
The course in five rules:
- Start with a simple solution: you do not always need an agent.
- An agent consists of a model, tools and instructions, and you can check each part separately.
- The agent takes knowledge from your documents, not from the model’s memory.
- Build the first agent on one task and test it on real examples.
- A person confirms important actions, and the company answers for the agent’s replies.
What should a manager know?
Send the course link to staff who will set tasks for agents or check their work. They finish the course in one sitting and take a test: answer 10 of 12 questions correctly. The employee sends the numbered certificate himself or herself. The site collects no names and does not know who passed.
The course does not teach programming. It gives the team a shared language: what a model, tools and instructions are, where an agent pays back and where a simple scenario is enough, and which actions you must not give an agent without a human check. With this language it is easier to discuss a project with a contractor and judge their offer.
How do you pick the first task for an agent?
The main mistake of a first project happens before the build: the agent gets the wrong task. The course teaches you to start with a simple solution and add complexity only when needed. If the steps are always the same, ordinary automation is more reliable and cheaper.
An agent is justified when the task has at least one of three features: decisions with exceptions, such as whether to approve a refund; heavy rules that are costly to keep in code, such as checking a supplier against dozens of conditions; and a lot of unstructured data, such as letters, scans and contracts.
Before the start, check the task with three questions. Can you describe the result in one sentence and check it? Are the data and access available? What happens if the agent errs, and who will notice? If there is no answer to the last question, do not give the task to an agent.
How do you handle data and agent rights?
An agent acts by itself, so its error costs more than a chatbot’s error. A chatbot is a program that exchanges messages with a person in a chat window. An agent can send a letter, change a deal or show something it should not. The course covers the rules of control: the agent gets only the needed tools and minimal rights, a person confirms payments and order cancellations, and a task the agent cannot handle goes to an employee.
A separate question concerns customer data. An agent connected to a CRM and to a foreign cloud modelcloud AI model. A cloud AI model is a neural network that runs on the servers of a provider. You use it through a website, an app or an API, and you pay for the volume of requests. sends it customer details in each request where they are needed for an answer. A cloud model is a neural network that runs on a provider’s servers; you use it through a site, an app or an API and pay for the volume of requests. So for work tasks with customer data the course advises an open-weight model on your own server. An open-weight model is a neural network whose parameter files you can download and run on your own hardware. Correspondence and documents then do not leave the company.
What does Russian law require?
Russia is not on the lists of countries where OpenAI services and the Anthropic API work, so a model for a Russian company is chosen with that in mind. Under 152-FZ, Russia’s personal data law, personal data of Russian citizens is recorded and stored in databases in Russia. An agent connected to a CRM and a foreign cloud model can break this rule. Since 30 May 2025, Law 420-FZ sets the fine for a personal data leak in Russia at $36,000–$180,000, depending on the volume. For a repeat leak the fine is 1–3% of yearly revenue. In 2025 the amount of data that left 150 Russian organisations for public AI services grew 30 times.
What do you do after the course?
The course ends with a plan for a first agent on one task. After that a company has three paths:
- An agent with clear steps and no access to customer data is built by an employee, by the instructions.
- An agent with access to a CRM, mail or a customer base is built with the person responsible for security.
- An agent for several departments or for your own product is made with a developer. The shared language from the course then helps to check the developer’s offer.