An always-on AI agent keeps running between messages. It is still there when nobody has written to it for six hours, and it does not need to be woken before it can act. "No server to manage" means something narrower than people assume: a machine still exists, you just do not patch it, restart it or hold the pager for it.
The short answer. There are three ways to get always-on. Rent a VPS and administer it yourself for about $6 per month plus your evenings. Put the agent in serverless functions and accept that it only exists while a request is in flight. Or buy a managed runtime, which is $49 per month on Sokko, where the machine is yours to use but not yours to maintain. If the agent has to be alive and you do not want to be an administrator, the third one is the only one that meets both conditions.
Table of Contents
The three ways to keep an agent alive
| Own VPS | Serverless functions | Managed runtime | |
|---|---|---|---|
| Who patches the OS | You | The provider | The provider |
| Does state survive a restart | Yes, if you built for it | No, unless you added a database | Yes, its files are kept |
| What you pay when idle | The full VPS price | Nothing | The plan price |
| What breaks at 2am | The process, and nobody notices | The request returns an error | Usually nothing; it restarts itself |
| Time to a running agent | An afternoon, then upkeep | A day, plus a database and a queue | Minutes |
| When it is the right answer | You want host-level control | The agent is request-scoped | The agent is a service |
The middle column is the one that catches people. Serverless is excellent for "answer this request and return", and it is the wrong shape for "watch this inbox and decide", because there is nothing running between invocations. An agent that has to notice something needs a process, not a function.
The four tests
Any platform can claim always-on. Four tests take an hour and settle it.
1. The sleep test. Message the agent, wait eight hours, message it again. A real resident agent answers with its context intact. A cron job with a chat window answers as if it has never met you, and a serverless agent answers but has forgotten the morning.
2. The restart test. Force a restart, or wait for the platform to do it. Then ask the agent to read a file it wrote before the restart. If the file is gone, the "persistent workspace" in the marketing copy is a cache.
3. The crash test. Break something on purpose: an invalid model key, a bad tool call. Does the agent come back on its own, and can you see what happened in the logs? Every platform promises uptime, and the difference between them is the quality of the evidence afterwards.
4. The deploy test. Change one line of the agent's configuration and ship it without opening a browser. If deploying requires clicking through a form, you cannot review the change, and the running agent will drift from whatever is in your repo.
Sokko is built around the first two: each agent gets its own isolated machine, its files persist across restarts, and it restarts itself if it crashes. That is also why a managed runtime costs more than a $6 VPS. You are buying the third column of the table above, not the hardware.
What it costs, including the line nobody budgets
The honest comparison has three lines, and only two of them are on an invoice.
| Own VPS | Managed runtime | |
|---|---|---|
| Hosting | $6 to $20 per month | $49 per month for two agents, $99 for four |
| Model calls | $10 to $60 per month, depending on volume | Same, inside credits or on your own keys |
| Your time | Setup afternoon, plus an evening per quarter, plus the morning it breaks | Minutes at setup, and reading logs when something looks wrong |
| Static IP, TLS, backups | Your problem | Included |
At an owner's rate, one broken morning a quarter is more expensive than the subscription difference for the whole year. The $6 VPS is a real option and a real bargain, and it is a bargain for people whose time is not on the invoice.
Two smaller lines worth asking about before you commit: EU data residency and where inference runs. Sokko offers EU residency and EU-hosted inference, which is the difference between passing a European security review and rescheduling it.
What you give up
Administration access. You get a terminal into the agent and readable configuration files. You do not get to choose the kernel, install a VPN daemon, or tune the network stack.
Exact pricing at fleet scale. Flat per-agent pricing is predictable and it is not infinitely elastic. Bakery covers four agents; a twenty-agent fleet is a conversation rather than a checkout.
Per-hour precision. A metered platform would let you pay for exactly the hours you use. A flat plan charges the same on a quiet Tuesday as on a busy one, which is a feature when you are budgeting and a cost when you are not.
A free tier. Sokko has none for hosting. The free plan runs a sandbox so you can try the product, and real hosting starts at $49.
Frequently asked questions
Is a laptop always-on if it never sleeps?
No. Closing the lid suspends the process, home IP addresses change, and a laptop that reboots after an update does not restart your agent unless you configured it to. The laptop failure modes are the reason people move to a server in the first place, and a managed runtime is the version of that move with the least work.
Does a managed agent sleep when it is idle?
It should not. Sleeping is a cost optimisation for the platform and a defect for you, because the wake-up latency lands on the first customer of the morning. Ask any vendor whether the agent's process stays resident between messages, and test it with the sleep test above.
What happens if the agent crashes at 3am?
On a managed runtime it should restart itself without you. Check the logs afterwards to see what happened, because a platform that restarts silently and keeps no evidence is hiding a problem rather than fixing it. On a VPS, nothing restarts unless you configured a supervisor.
Can I move the agent later?
Ask this before you sign. What you want is the ability to walk away with the agent's configuration, its instructions and its notes. If those only exist inside a vendor's database, you are renting the agent rather than owning it, and that is a decision worth making deliberately.
Do I need a static IP?
Only if something you connect to allowlists by address. Most integrations authenticate with a token instead, which is why managed hosting can sit behind shared infrastructure without breaking your tools.
Where Sokko is the wrong choice
You want host-level control. Run the VPS. The VPS hosting guide covers the setup and the maintenance tax honestly.
The agent is request-scoped. A webhook that answers and returns belongs in a function, and paying for a resident machine is waste.
You have a platform team. If you already run Kubernetes, running the agent there is not a burden, and it keeps everything in one place.
More than four agents on a card. That needs a commercial conversation, and self-hosting may genuinely be cheaper.
If you are still deciding between always-on and session-based designs, the comparison of the two is the closer read. For the full platform shortlist, start with the ten-platform comparison.