The difference between chatting with a neural network and an AI agent is simple: you explain to a chat what to do and get text back; you explain to an agent what you want, and it breaks the task into steps, uses tools—files, browser, email, terminal—and drives the job to a result. In 2026, agents are no longer conference demos: Anthropic and OpenAI have released official guides on building them, and ready-made skill sets let you assemble an assistant without a single line of code. We break down how to do it, what it can do, and where it's best not to let it near.
What an agent can do that a chat cannot
An agent is the same language model, but with three additions:
- Tools. The model can call functions: read a file, run a command, make an internet request, send a message. This is called tool use.
- Loop. The agent doesn't answer once; it works in a loop of "thought—action—observed result—thought again" until the task is done or hits a limit.
- Memory. The agent remembers context between runs: what has already been done, what your preferences are, what constraints exist.
In practice, it looks like this. You write to a chat, "write a sales report for the week," and get a template with made-up numbers. You write the same to an agent, and it opens the sales spreadsheet, calculates metrics, compares with last week, writes a report with real data, and sends it to the right chat.
Which tasks to entrust to an agent first
Good candidates for automation are tasks that repeat, have a clear completion criterion, and where the cost of an error is low. From Anthropic's guide on agents, which we wrote about in our channel, a typical first agent can:
- daily monitor GitHub and Slack (or their analogues), find issues, and suggest improvements;
- track new messages and changes in projects;
- independently close simple tasks and send reports;
- remember context, take your recommendations into account, and stay within the token limit.
If we translate this to office work, we get the following list of starter scenarios:
- Sorting incoming email: categorize emails by importance, draft replies to typical requests, and put anything requiring your decision into a separate list.
- Morning digest: collect changes from work chats, the task tracker, and the calendar into one short digest.
- Deadline control: check tasks with approaching deadlines, remind assignees, and flag overdue ones.
- Weekly report: collect data from spreadsheets and systems and compile it into a document using a template.
- Competitor and niche monitoring: once a day, review specified sites and channels and note anything new.
What you shouldn't entrust to an agent yet: sending money, deleting data, public replies on your behalf without review, or any action that can't be undone. The rule: the agent prepares, the human confirms—until you've gathered statistics on its mistakes.
Which tool to choose
You don't need to write code from scratch to build an agent. In 2026, there are four mature options, and the choice depends on where you work and what you're willing to pay for.
Claude Code
Claude Code is Anthropic's agent that works in the terminal and in code editors. It was originally created for programmers, but it handles office tasks well too: it can work with files, run scripts, go online, and connect external services. Its main advantage is the system of skills and subagents: you describe a role in a text file, and the agent performs it. To get started, there's a free training course that covers everything from the basics to managing hundreds of agents. The cost is a subscription from $20 per month.
Codex CLI and Gemini CLI
Codex CLI from OpenAI and Gemini CLI from Google are analogues of Claude Code from the other two labs. Codex comes with a ChatGPT subscription, and Gemini CLI has a generous free tier, making it the best option for trying agents for free. Functionally, all three tools are similar; the difference is in the models under the hood and the extension ecosystem.
DeepSeek desktop app
In October, DeepSeek released an open-source app for Windows and macOS. It works with files, code, and the terminal on your computer, searches the internet, and consists of plugins—missing functions are added as in a constructor. The app itself is free; you only pay for model tokens. A good choice if openness and the ability to connect any model, including a local one, matter to you.
Visual builders
If the terminal intimidates you, there are platforms with a visual scenario editor: Botpress, Voiceflow, and similar. They are convenient for chatbots and customer support, but for agents working with your files and systems, terminal tools are more flexible. You can compare all the options in the catalog of development tools.
Step by step: your first agent in one evening
Let's use the example of an agent for a morning digest that collects changes from work sources into one digest. The instructions work for any of the terminal tools above; only the syntax for connecting services differs.
Step 1. Describe the role and goal
Create a text file with instructions for the agent. This is the most important step—80% of the result depends on it. A good instruction contains four blocks:
Role: you are my morning digest assistant.
Goal: every morning by 9:00, collect everything that changed in the last
24 hours in tasks, work chats, and the calendar, and send me
a short digest.
Result format:
- no more than 10 items, each one line;
- first what requires my action today, then the rest;
- at the end—three meetings for today with times.
Constraints:
- do not send anything on my behalf and do not delete anything;
- if a source is unavailable, write one line about it and continue;
- do not spend more than 50 thousand tokens per run.
Note the constraints: an explicit ban on actions on your behalf and a token limit protect against the two most common problems—the agent acting on its own initiative and an unexpected API bill.
Step 2. Connect sources
An agent needs access to data. In terminal tools, this is done by connecting external services (in the Claude ecosystem and others, these are called MCP servers): for task trackers, email, calendars, and messengers, there are ready-made connections that are enabled with one command from the documentation. To start, connect one or two sources, not all at once.
Step 3. Run it manually and watch the process
Do your first runs in interactive mode and read what the agent does at each step. You'll quickly see where it misunderstands the task: for example, it considers everything "important" or gets stuck on an unavailable source. Edit the instruction after each run. Usually three to five iterations are enough for the digest to become useful.
Step 4. Add memory
To make the agent take your preferences into account, give it a notes file that it reads at the start and can add to. For example: "the user asked not to include notifications from the CI bot" or "don't show meetings marked optional." This way the agent gets better from run to run without rewriting the instruction.
Step 5. Schedule it
When you're satisfied with the result, run the agent automatically: with the operating system's task scheduler, cron on a server, or the tool's built-in schedule. Send the result where you'll definitely read it—most often to personal messages in Telegram via a bot.
Ready-made skills instead of instructions from scratch
You don't have to write roles from scratch. In October, developers published a set of 338 ready-made skills for agents—from CEO, CTO, and CFO to marketer, lawyer, analyst, and personal assistant. Each skill is a well-developed instruction for a specific role: an agent with the analyst role can conduct market analysis, with the SEO specialist role—audit a website, with the financier role—calculate budgets. Take a suitable role as a basis and adapt it to your context—it's faster and better than starting from a blank page.
Which model to choose for an agent
An agent makes dozens of model calls per run, so the price of tokens matters more here than in a chat. A reasonable strategy for October 2026:
- For routine (sorting, digests, reminders)—fast budget models. Claude Haiku 5.5 shows 72% in computer-use tasks at a price 10 times lower than the senior models; for OpenAI, GPT-6 Luna plays a similar role.
- For reasoning tasks (analysis, code, decisions on ambiguous criteria)—the mid tier: Claude Sonnet 5.5 or GPT-6.1 Sol.
- For confidential data—a local open model via Ollama; we covered how to run it in a separate guide.
A good technique is to combine: a cheap model does the rough work, and an expensive one is connected only for complex steps. This is roughly what OpenAI's guide on distributing tasks between models describes.
Typical beginner mistakes
Too general an instruction. "Help me with work" doesn't work. An agent needs a specific goal, result format, and completion criterion.
Too many tools at once. By connecting ten services, you'll get an agent that gets lost in the choices. Start with one or two and add as needed.
No limits. Without a limit on tokens and number of steps, an agent can get stuck in a loop and spend a month's budget overnight. Always set the limit explicitly.
Trust without verification. For the first few weeks, read everything the agent does, especially if it sends or changes something. Language models hallucinate, and an agent can confidently report something it didn't do.
Expecting perfection on the first try. An agent is a setup process, not an out-of-the-box product. Five iterations of the instruction is normal.
Takeaways
The first useful agent can be built in one evening: describe the role and constraints, connect one or two data sources, run it manually several times, add a memory file, and schedule it. Start with a task where a mistake isn't scary—a morning digest or email sorting—and expand the agent's authority as trust grows. Ready-made skills and official guides from Anthropic and OpenAI greatly shorten the path: today this is no longer research work but tool configuration.
Agent tools are updated every week: new skills, connections, and models with a better price-quality ratio. We publish the most important of this in our channel on the day it comes out.