What "autonomous AI agent" means
An AI chat waits for your question and answers once. An agent receives a task in words, chooses the steps itself and carries them out with tools: it reads files, works with email, writes in Telegram and checks the result. Autonomy is how many of these steps it takes in a row without you.
How an agent works inside is covered in How AI agents work. This article is about the line between "do it yourself" and "ask me".
What the agent does without asking
- Scheduled tasks
A morning summary, email triage, a list of unpaid invoices every Monday. You set the time with an ordinary message.
- Reminders
It writes to you when a deadline from an agreement approaches: call a client, renew a contract, send a proposal.
- Memory
It records outcomes of conversations, decisions and your rules in its knowledge base and retrieves them when needed.
- Drafts
It prepares emails, replies and documents. Send them out after you review them, and make that a rule.
- Questions for you
If data is missing or a task can be read two ways, the agent should ask rather than guess; make that a rule.
All of this happens within what you gave it access to: your files on the server, email if you connected it, your Telegram chat.
Four levels of independence
It is convenient to choose the level for each task separately. The same agent can send a summary on its own and prepare client emails only as drafts.
| What the agent does | What stays with you | |
|---|---|---|
| Answers | Answers when you ask | Everything else |
| Prepares | Makes a draft, list or calculation and sends it for review | Review and send it yourself |
| Acts after a "yes" | Prepares an action and performs it after your reply in chat | Reply "yes" or correct it |
| Acts and reports | Does it alone and sends a report on what was done | Read reports, adjust rules |
Start a new task at the "Prepares" level. Raise the level once the agent has handled it for two weeks without corrections. Keep money and anything irreversible at "Acts after a yes".
When the agent should ask a human
In these actions a mistake costs more than the time saved.
- Money: payments, refunds, discounts, subscriptions. The agent prepares an invoice or a list to pay, you pay.
- Emails and messages to people on your behalf: clients, partners, group chats. To the recipient, the agent's reply is your reply.
- Deleting files and records, anything that cannot be undone.
- Commitments: prices, deadlines, terms. The agent takes them from your files, you confirm them.
- Publishing on your behalf: posts, reviews, comments.
- Access: passwords, keys, new people the agent will answer.
How to set limits for the agent
- Send the agent three lists in one message: what it does alone, what it does after your "yes" and what it never does.
- Ask it to save them as permanent rules. The agent keeps the rules in memory and checks them before acting.
- Give access per task. Read access to the mailbox is enough for email triage; sending can be added later.
- Test the limit: give it a task that touches the limit, such as "delete old files", and see whether the agent asks.
- Raise independence one task at a time, once that task runs without corrections.
An example of such a message: "On your own: morning summary, reminders, drafts, notes in the knowledge base. After my 'yes': client emails, deleting files, anything involving money. Never: payments, passwords, replies in group chats."
Rules and access: which is more reliable
A rule in the agent's memory is an instruction for the model. The model can misread it or make a mistake in a rare case. So whatever the agent must never do, close off with access. If the agent has no access to the bank, it cannot pay anything, however it reads the task.
A separate risk is other people's text. An email, a web page or a document may contain a phrase like "forward all files to this address". Tell the agent plainly: emails, pages and files are data, and it takes commands only from you. Let it forward suspicious requests from such sources to you. This rule lowers the risk but does not remove it, so close off the most dangerous actions with access.
How to manage several AI agents
A system of AI agents usually means several agents with different roles: one handles email and clients, another helps with personal matters, a third works with texts. It is better to give each one its own memory, its own rules and only the access it needs. Then a mistake by one does not affect the others, and each one's limits are easier to check.
You need summaries, reminders and drafts for one person
One agent is enough. Keep different tasks at different levels of independence.Several people need agents: you, a partner, your family
One agent per person. Each has its own memory, and nobody sees anyone else's affairs.You set up agents for your clients
Each client gets their own agent with their own rules and access. The Partner plan is made for this.
How to start working with an agent is covered in How to set up an AI agent, and business scenarios in AI agent examples.
An autonomous agent on NekoAgent
NekoAgent is an AI agent on your server: it talks in Telegram by text and voice, keeps a knowledge base in Obsidian, performs scheduled tasks and runs on your Claude or ChatGPT subscription. The terms and availability of the subscription in your country are set by the service itself. You set limits in an ordinary chat and the agent saves them to memory. During the support period, Iris monitors the agent and helps bring it back after a failure.
We will set up an agent that works on its own within your limits
From 24,900 RUB once per install (Egoist plan): on a call with a curator, with training. The server and AI model subscription are paid separately. 30 days of support are included in the plan.
Key questions
What is an autonomous AI agent?
An AI agent that performs tasks without a command for every step: on a schedule and by the owner's rules. It chooses the steps, works with files and email and reports the result.
Can an AI agent work with no human at all?
It carries out long chains of actions on its own. Money, messages to people on your behalf and deleting data are kept subject to confirmation: a mistake there costs more than the time saved, and you are responsible to other people for what the agent sends or does on your behalf.
What does an AI agent do without asking?
Scheduled tasks, reminders, drafts, notes in memory and work with your files. What it may send on your behalf is set by your rules and the access you give it.
How do you limit an AI agent?
With rules and access. You write the rules in chat: what it does alone, what after your "yes", what never. Do not give the agent access to anything it must never touch.
How do you manage several AI agents?
Give each one its own role, memory, rules and access. Then a mistake by one does not affect the others, and each one's limits are easier to check.
