Which AI agent is best: types of agents and how to choose yours

What kinds of AI agents there are

“AI agent” now means very different things: from a chain of blocks in a cloud service to a program running around the clock on your server. ChatGPT and Claude in a browser do not count: they are assistants that answer in a chat while you talk to them. How they differ from agents is covered in AI agent vs AI assistant.

An agent works on its own: it remembers people and agreements, runs tasks on a schedule and messages you first. Such agents fall into four types.

1. Cloud builders

Automation services such as n8n and Make and no-code bot platforms. You assemble the agent with the mouse from ready-made blocks, no server of your own needed. It can do what the blocks allow, and data and memory are kept by the platform. You usually pay per operation, and the bill grows with the load.

2. Open-source agents: OpenClaw and Hermes Agent

Strong open-source projects with large communities, and the software is free. OpenClaw is installed in the terminal with the openclaw onboard wizard, connects to more than ten messengers and takes skills from the ClawHub catalog. Hermes Agent from Nous Research is installed with a script and hermes setup, supports more than 200 models and learns from its own experience.

Server, updates, backups and security are on the owner, as the OpenClaw docs state plainly. The Claude subscription rules changed during 2026: in spring, from April 4, third-party harnesses required extra payment or an API key. As of October 28, 2026, the Anthropic help article says that third-party apps signed in with a Claude plan draw from the subscription limits. Hermes has its own terms: according to its docs, signing in via Anthropic works only on the Max plan and is billed as extra usage.

3. Your own agent on Claude Code or Codex

Claude Code from Anthropic and Codex from OpenAI are harnesses that give the model files and the server terminal. A regular subscription covers them within the plan's limits: Claude Pro or Max for Claude Code, ChatGPT for Codex. But this is a starting kit: memory, scheduling, restarts after failures and the Telegram link you have to write yourself. This path is for those who code or are ready to spend a long time figuring it out.

4. Ready-made turnkey agents

The agent is installed on your server already set up, and the provider gives support under the terms of its offer. That is how NekoAgent works: inside are Claude Code or Codex, so a regular Claude or ChatGPT subscription works. Memory is an Obsidian knowledge base on your server, and you talk to it in Telegram by text and voice. The Iris support service looks after the agent.

Five questions to choose an agent by

An agent's name tells you less than the answers to five questions. Ask them about any option you are considering.

• Where the data lives

  On your server or with the platform. This decides who sees the chats and the knowledge base and whether you can take them with you.

• Who fixes and updates it

  You, the platform or a support service. Updates sometimes break things, and someone has to fix them.

• Monthly cost

  Add everything up: server, model, support. Free software does not make the agent free.

• Is there Telegram

  Can you set tasks from your phone, by text and voice, and does the agent message you itself.

• Which subscription works

  Does your Claude or ChatGPT subscription work, or will you pay per token via the API.

We put the answers for the four types into a table.

Why we do not publish a “best of” ranking

Rankings compare programs. But an agent at work is the program together with a server, a model and the person who looks after it. The same OpenClaw lives very differently with someone at home in the terminal and with a shop owner who has never opened one.

So the best agent is the one whose answers to the five questions suit you: data sits where you feel safe, there is someone to fix it, the monthly bill is clear, and setting tasks is convenient.

How to choose

• One narrow task: sort requests into a spreadsheet, answer a typical question

  →A cloud builder

• You are comfortable in a Linux terminal and will look after the server yourself

  →OpenClaw, Hermes or your own agent on Claude Code

• You need an everyday assistant and have no time to fix it

  →A ready-made turnkey agent

What to ask before buying a ready-made agent

If you choose a turnkey agent from any provider, us included, ask these questions before paying.

1. Whose server will the agent run on, and can I take my data if I leave?
2. Whose AI subscription is it: mine or yours?
3. Who brings the agent back after a failure, and what does support cost after the first month?
4. Are there backups, and where are they kept?
5. How will I set tasks: in Telegram, by voice, on a schedule?

Our answers. The agent runs on your VPS with 2 GB of RAM or more, about 1,500 RUB a month. The data stays on your server; encrypted backups are kept by us in Russia, and the support service accesses the server with a key you can remove at any time. The Claude or ChatGPT subscription is yours, and you pay the service directly. The Iris support service helps bring the agent back after failures and makes nightly encrypted backups; support is free for the first 30 days, then 750 RUB a month for every 5 agents. You set tasks in Telegram by text or voice, and the agent does regular ones on a schedule.

Installation paths with timelines are covered in How to create your own AI agent, and the costs of free options 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.

Plans and prices Ask a question

Key questions