> ## Documentation Index
> Fetch the complete documentation index at: https://ngrok.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# How ngrok Agents Work

> What a session is, how agents create endpoints, and what happens when an agent stops or loses its connection.

Every kind of agent does the same three things: it opens a session with ngrok, it creates endpoints on that session, and it carries traffic back to your upstream service.
Those three pieces explain most of the agent's behavior, including the parts that surprise people.

## The session

When an agent starts, it dials out to ngrok and establishes a long-lived TLS connection on port 443.
That connection is the **session**, and it's the unit ngrok counts: the Agents page in your dashboard lists sessions, whether they were opened by the standalone agent, an application using an Agent SDK, a Kubernetes Operator, or a reverse SSH client.

The session is outbound only.
Nothing listens for inbound connections on your machine, which is why agents work behind NAT, corporate firewalls, and networks you don't control.
It's also why there's nothing to port forward and no inbound firewall rule to add.

The session is also the handle you manage an agent by.
Because ngrok holds one end of it, you can stop, restart, or upgrade an agent from the dashboard or the [Tunnel Sessions API](/docs/api-reference/tunnelsessions/list/) even when you can't reach the machine itself.

The agent keeps the session alive with periodic heartbeats and reconnects on its own if the network drops.
[Connectivity](/docs/gateway/agent/connectivity) covers that machinery, and how to diagnose it when the session won't establish.

## Endpoints on that session

Once the session is up, the agent asks ngrok to create **endpoints** on it.
An endpoint is the URL traffic arrives at, and the agent tells ngrok which upstream service to forward it to.

Endpoints an agent creates are [Agent Endpoints](/docs/gateway/endpoints/agent-endpoints), and they differ from [Cloud Endpoints](/docs/gateway/endpoints/cloud-endpoints/) in ways that follow directly from living on a session:

* **They last as long as the session does.** Stop the agent and the endpoint stops answering.
  A Cloud Endpoint is configured centrally and stays up whether or not any process of yours is running.
* **The agent is authoritative.** The agent that created an endpoint owns its configuration.
  The dashboard and API give you a read-only view.
* **Forwarding is implicit.** Traffic goes to the agent that created the endpoint, because that's the connection it arrived on.
* **Traffic Policy is optional.** Anything a policy doesn't terminate goes to the agent.

One agent can hold many endpoints on a single session, which is why [a configuration file](/docs/gateway/agent/config/) is the usual way to run more than one service.

## When an agent stops

Endpoints live on the session, so they go offline when the agent does.
That's the single most important consequence of the model, and most agent operations exist to manage it:

* Run the agent [as a native OS service](/docs/gateway/agent/service) so it starts on boot and restarts after a crash.
* Give an endpoint a [reserved domain](/docs/gateway/domains/) so its URL survives a restart instead of being reassigned.
* Put two agents behind [endpoint pooling](/docs/gateway/endpoints/endpoint-pooling/) so one can fail without taking the endpoint down.

If you need the URL to stay up independently of any process you run, that's a Cloud Endpoint, not an agent.

## What the agent doesn't do

The agent is a connection and forwarding component, not a proxy you configure with rules.
Authentication, routing, rewriting, and rate limiting are described by [Traffic Policy](/docs/gateway/traffic-policy/), which ngrok applies in its own network before traffic reaches your agent.
This is why the same policy works identically no matter which kind of agent created the endpoint.