Can I deploy an AI agent on serverless functions?

Short, stateless steps, yes. Long-running tasks, schedules, chat bridges and memory fit badly because of execution limits and cold starts. An always-on instance avoids all of it.

How much memory does an AI agent need?

Around 4 GB for a lightweight assistant with chat bridges, 8 GB for frameworks that bundle a database, workers and a vector store. The model itself runs at your provider unless you host one locally.

Do you provide the model?

No. You bring an API key from your provider, billed to you at list price, or pair the agent with a local model plan.

Can the agent run on a schedule?

Yes. The instance runs continuously, so scheduled tasks fire whether or not anyone is logged in.

Is my agent isolated from other customers?

Yes. Each agent runs on its own single-tenant instance with its own storage and secrets.

AI Agent Hosting — Deploy Agents That Stay Running

An agent that works on a laptop is a demo. An agent that answers at 3am, runs its Monday report on its own and still remembers last week’s conversation is a service — and services need somewhere to live. This is what that somewhere has to provide.

Agents are not ordinary web apps

A typical web app waits for a request, answers it in a few hundred milliseconds and forgets it. Agents break every part of that. A single task can run for minutes while the model plans, calls tools and retries. Many agents are not triggered by a request at all but by a schedule, a message in a chat app, or an event from another system. And the useful ones carry state: memory of past conversations, files they are working on, sessions with the services they are connected to.

So the question “where do I deploy my agent” is really about four things: something that stays up, somewhere state survives, a safe place for credentials, and a way to see what the agent did.

Why serverless is the wrong shape

Function platforms are built for short, stateless work. They cap execution time, they shut down between calls, and they start cold. An agent halfway through a ten-step task that hits a time limit does not fail cleanly — it stops mid-action, and you find out when someone notices the task was half done.

Chat bridges make it worse. An assistant paired to WhatsApp, Telegram or Slack needs a long-lived connection and a pairing that survives restarts. That is a process that runs continuously, which is exactly what serverless is designed not to be.

What an agent host needs

An always-on process with restart on failure. If the agent crashes at 2am, it should be back before anyone notices, without someone logging in.

Persistent storage. Memory, workspaces, pairings and job queues must survive restarts and updates. Losing them means re-pairing every bridge and an agent that forgets everyone.

Isolated secrets. An agent holds model keys and often credentials to your mail, calendar, CRM or database. Those belong on a single-tenant instance, not in a shared environment or a config file on someone’s laptop.

Outbound network, controlled inbound. Agents reach out to APIs and websites; very few need to accept traffic from the internet beyond a webhook or the admin interface, which should sit behind TLS and authentication.

Logs and backups. When an agent does something surprising, you need to see what it did and why, and be able to roll back what it changed.

The agents we host, and what each is for

OpenClaw — a personal or team assistant that lives in your chat apps, runs skills and scheduled tasks. From $9.99 on 4 GB; the 8 GB plan at $16.99 adds headroom for browser automation.

Letta — agents with long-term memory as a first-class feature, for assistants that need to remember users and context across months. From $11.99.

OpenHands — a software-engineering agent that reads a repository, runs commands and proposes changes. From $18.99 on 8 GB.

Dify and Flowise — visual builders for agent workflows with tools and retrieval, when the logic should be editable by people who do not write code. From $18.99 and $11.99.

Browser Use — agents that operate a real browser for sites without an API. From $18.99.

The full catalogue lists the rest. Each runs on its own dedicated instance, not a shared container.

Running an agent safely

Scope every credential. An agent that only needs to read a mailbox should not hold a key that can send from it. Give it the narrowest access that does the job.

Keep a human on destructive actions. Deleting records, sending money, emailing customers — have the agent draft and a person approve until you have watched it do the task correctly many times.

Cap the model spend. A loop that retries forever is also a bill that grows forever. Routing the agent through LiteLLM gives it a budget it cannot exceed.

Pin the version. Agent frameworks change weekly. We update deliberately after a release has been reviewed, not automatically underneath a running agent.

What it costs

Agents themselves are light. The gateway process and bridges of an assistant like OpenClaw fit in 4 GB; heavier frameworks with a database, a worker and a vector store need 8 GB. That puts most agent hosting between $9.99 and $18.99 a month, flat.

The model is billed separately by your provider, on your own key, and it is usually the larger number for an active agent. Use a frontier model for planning and a cheaper or local model for routine steps; our guide to cutting LLM costs covers the routing patterns that make the biggest difference.

Getting an agent live

Pick the framework and plan. We provision a dedicated instance with persistent storage, TLS on your domain or ours, and the version pinned. You add your model key and connect the services the agent should reach. For chat-based assistants, plan to be at a keyboard for the first pairing — it is interactive by design.

After that, uptime, restarts, updates and nightly off-box backups are ours; what the agent does is yours.

AI agent hosting: where an agent should actually run