What an AI agent is, in plain words
A regular AI chat answers with words. You ask, it writes an answer, and the rest is up to you: copy the text, send the email, set the reminder.
An agent sees the job through. You can tell it “send me an email summary every morning” or “remind the client about the payment on Friday”, and it will do it on its own, at the right time. For that the model needs hands, memory and a clock, and that is what the agent's design provides.
What an agent is made of
Take an agent apart and you find six parts. Each has its own job.
- Model
Thinks and makes decisions: Claude from Anthropic or GPT from OpenAI. On its own a model only reads text and writes text.
- Harness
The program around the model: passes it the task, carries out its decisions and returns the result. Examples are Claude Code, Codex, OpenClaw and Hermes Agent.
- Tools
What the agent acts with: files, email, calendar, the server terminal. Each tool is an action the harness performs when the model asks.
- Memory
Notes the agent keeps itself: people, agreements, decisions, task results. It rereads them before working and does not start from scratch every time.
- Schedule
The agent's alarm clock. At the right time the harness wakes the model and gives it a task: a morning summary, an inbox check, a reminder.
- Channel
Usually a Telegram bot. You write to the agent in text or voice, it answers and messages you itself when it has something to say.
The work loop: task, plan, action, check
The heart of an agent is the loop. The model works in steps and repeats them until the task is done.
- Task. It comes from you in Telegram or from the schedule. For example: “find which clients have not replied to emails this week”.
- Plan. The model decides which steps are needed: open the inbox, find unanswered emails, check them against the client notes.
- Action. The model asks the harness to perform a step, such as reading the inbox. The harness does it and returns the result.
- Check. The model looks at the result: did it work, is anything missing. If needed, it changes the plan and takes the next step.
- Result. When the task is done, the agent sends you the answer and writes what it did into memory.
One task can take many rounds. The agent decides when the task is done and when it is time to ask you.
How an agent remembers
A model on its own does not remember past conversations: each time it sees only what was passed in the request. So an agent's memory is external: notes it writes and rereads.
Good memory works like a notebook: separate notes about people, projects and decisions. Before a task the agent finds the right ones, after it adds what is new. In NekoAgent this is an Obsidian knowledge base on the client's server: you can open and edit the notes by hand, and they survive restarts.
How an agent differs from an AI chat
| AI chat in a browser | AI agent | |
|---|---|---|
| When it works | While the chat is open | Around the clock on a server |
| Actions | Writes text, you do the rest | Works with files, email and the server itself |
| Initiative | Answers questions | Messages you on schedule and reminds you of tasks |
| Memory | Whatever the service saved | Its own knowledge base you can open and edit |
| Where it lives | With the service | On a server, for example yours |
Where an agent goes wrong and who keeps an eye on it
An agent makes the same mistakes as the model inside it: it may misread the task or pick the wrong step. The check in the loop catches some errors, not all. Three habits help.
- Make the task specific: what to do, by when and what counts as done.
- Keep important actions, such as emails to clients, payments and deleting files, for your confirmation.
- Correct the agent in words: “don't do it like that, do it like this”. A good agent writes the correction into memory and applies it next time.
An agent is also a program on a server, and it can crash or go out of date. In NekoAgent the Iris support service handles this during paid support: 24/7 monitoring, help restarting a crashed agent, nightly encrypted backups and tested core updates you install with one button.
How to build your own agent is covered in How to create your own AI agent: DIY or turnkey, and what can be free in it in Free AI agents.
Let's set up your agent together
Plans from 24,900 RUB for three installs. The first one is on a call with a curator, with training.
Key questions
What is an AI agent in plain words?
A language model that has been given tools, memory and a schedule. It carries out tasks itself: works with email and files, reminds you of things and messages you first.
How does an AI agent work?
In a loop: it gets a task, makes a plan, takes a step through its tools, checks the result and repeats until the task is done. It reports the result to you and writes it into memory.
Which AI model does an agent need?
A model that handles tools well, such as Claude from Anthropic or GPT from OpenAI. NekoAgent runs on Claude Code or Codex, covered by a Claude or ChatGPT subscription.
How is an AI agent different from a chatbot?
A chatbot answers while you talk to it. An agent lives on a server, remembers people and decisions, works on a schedule and carries out tasks itself.
Does the agent do everything itself?
It carries out the tasks, and the decisions stay with you. Important actions such as client emails or payments are best left for your confirmation, and mistakes are corrected in words: the agent remembers corrections.
