Where Does Your AI Agent Actually Live Between Tasks?
You ask an AI something in ChatGPT. It answers. You close the tab. Everything is gone — the context, the plan it was building, the files it referenced. Next time you open a new chat, it starts from zero.
That's fine for a quick question. It's completely broken for an autonomous agent.
An autonomous agent is supposed to do things over time: check your inbox every morning, run a market scan every hour, remember what you told it last Tuesday, pick up where it left off after a weekend. For that to work, the agent can't live inside a chat tab. It needs a proper home.
This post explains what "hosting an AI agent" actually means under the hood — so you can make a real choice about how to set yours up.
The difference between a chat session and a hosted agent
A chat session is stateless by design. The model receives your message plus whatever chat history fits in its context window, generates a reply, and exits. The moment the response is sent, no computation remains running anywhere. There's no "it" left.
A hosted agent is different in three fundamental ways:
1. A persistent process that keeps running
Somewhere, a process — a piece of software — is alive on a machine. It's looping, listening for events, and waiting to act. When a trigger fires (a scheduled time, an incoming message, a webhook), the process wakes up, calls the LLM, takes action, and goes back to waiting. When the action is done, the process doesn't die. It stays.
2. Storage that survives restarts
The agent needs somewhere to write things down. Conversation history. The result of the last scan. A list of tasks in progress. That state lives in files or a database on a disk, not in RAM. If the machine restarts or the process crashes and relaunches, the agent reads the state back from disk and picks up where it left off.
3. A scheduler that fires tasks on time
"Check my email every morning at 8 AM" requires something to actually wake the agent up at 8 AM. That's a scheduler — either a cron job, a built-in task loop, or an event bus. Without one, your agent only runs when you manually send it a message.
Put all three together and you have an agent with a real home. Take any one away and you have something that only pretends to be autonomous.
What breaks without each piece
It helps to think about what goes wrong when you're missing each component.
| Missing piece | What breaks |
|---|---|
| Persistent process | Agent only runs when you're actively talking to it |
| Persistent storage | Forgets every conversation; can't track ongoing tasks |
| Scheduler | You become the trigger; nothing runs unless you initiate it |
Most "AI assistant" demos you see are missing all three. They're impressive for a five-minute video, and useless for anything you actually want to automate.
The self-hosting reality check
You can build all of this yourself. Here's what that looks like:
- Rent a VPS ($5–20/month depending on specs)
- Install your agent framework and its dependencies
- Write or configure the agent's persistent loop
- Set up a process supervisor (
systemd,supervisord,pm2) so it restarts on crash - Set up a cron job or in-process scheduler for timed tasks
- Mount a persistent volume or configure a database for state storage
- Get an API key from an LLM provider, handle billing, monitor usage
- Set up SSL if the agent needs a public endpoint (Telegram bots, webhooks)
- Debug why the process crashed at 3 AM
For a backend engineer, this is maybe half a day. For everyone else, it's a project that eats the weekend and still doesn't feel stable.
What a "persistent home" concretely looks like in practice
Let's make this concrete with a real agent type: a research and monitoring assistant that you want to:
- Check a set of URLs or RSS feeds every four hours
- Summarize anything new into a short digest
- Remember what it already summarized so it doesn't repeat itself
- Send the digest to you over Telegram
For this to work:
- The process needs to be alive around the clock, not just when you're at your computer
- The storage needs a file or table that tracks "already seen" items — otherwise every run re-summarizes everything
- The scheduler needs to wake the process every four hours and kick off the scan
- The Telegram integration needs an always-on bot that can push messages to you (not just respond to your inputs)
Every one of these is a moving part. A true autonomous agent hosting solution wires them all together before you even log in.
How Hermes and OpenClaw handle this
AgentRoost's Hermes and OpenClaw frameworks are built specifically around these three requirements — persistence, storage, and scheduling — for people who want an autonomous agent without the infrastructure project.
Hermes is a persistent AI assistant framework designed for things like the research helper above: scheduled tasks, long-running context, and multi-step workflows. It runs as an always-on process with conversation and task state stored to disk. You trigger it through an auto-provisioned Telegram bot — you just /start it. Scheduled tasks are configured inside the framework itself; you don't touch cron. The agent runs whether or not you're at your keyboard.
OpenClaw is a more focused persistent personal assistant: always-on, file and conversation state that survives restarts, delivered through a private Telegram bot. You talk to it; it remembers. You close the app; it's still there on the server, ready for the next message.
In both cases the LLM calls are powered by included AI credits — you don't register with OpenAI, Anthropic, or any LLM provider yourself. The credits are part of the subscription. This matters for autonomous agents because unattended scheduled tasks make API calls without you watching a usage dashboard. With bring-your-own-key hosting you can wake up to surprise charges. With AgentRoost, the cost is flat and known upfront.
How to get started on AgentRoost
The actual flow takes about two minutes:
- Sign up at agentroost.app (email/password, or Google, Microsoft, or Discord OAuth)
- Go to Agents and pick Hermes or OpenClaw depending on your use case
- Name your agent
- Open the AgentRoost manager bot in Telegram and
/startyour agent - The agent is live — it has its persistent home, its scheduler, its storage, and its included AI credits
No server provisioning. No systemd config. No API key registration. No SSL setup.
Plans start at $19.99/month, which includes the compute and the AI credits. There's a 14-day money-back guarantee, billed monthly, cancel anytime.
Compare plans and see what's included →
Choosing the right setup
If you're comfortable with Linux administration and want maximum control, self-hosting is a valid choice. Set up systemd, handle your own LLM billing, and own every moving part.
If you want an agent that's actually running tonight without a weekend of DevOps, AgentRoost is built for that. The infrastructure decisions — process supervision, storage, scheduling, LLM billing — are made for you, so you can spend the time on what the agent should actually do.
The question isn't which approach is "better" in the abstract. It's whether you want to maintain the home or just live in it.
Frequently asked questions
Do I need an API key for the LLM if I host on AgentRoost?
No. AI/LLM credits are included in every AgentRoost subscription. The agent's scheduled tasks and conversations use those included credits — you don't register with OpenAI, Anthropic, or any other provider yourself.
What happens to my agent's memory if AgentRoost restarts the container?
Conversation history and task state are written to persistent storage (disk), not held in RAM. When the process restarts after any maintenance or crash, the agent reads its state back from disk and continues where it left off.
Can my agent actually run tasks while I'm asleep or offline?
Yes. The agent process runs on AgentRoost's servers continuously. Scheduled tasks fire based on the schedule you configure, regardless of whether you have the app open or any device switched on.
How is this different from just using ChatGPT or Claude in a browser?
Browser-based chat is stateless — every session starts fresh, nothing runs between your messages, and there's no scheduler. A hosted agent on AgentRoost has a running process that persists state across days and fires tasks on a schedule without you initiating anything.
Can I cancel if it's not for me?
Yes. Plans are billed monthly with no annual commitment required. There is also a 14-day money-back guarantee if you decide the product isn't the right fit.