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

# Agent sessions

> Chat with the agent, give it the right credentials, reference your resources, and watch it work.

A **session** is a conversation with the agent. You describe what you want in
plain language and the agent investigates and acts against your connected
infrastructure, streaming its progress as it goes. The workspace is tab-based, so
you can run several sessions side by side.

## Starting a session

Open the **Agent** and start a new session. Each session keeps its own
conversation history, so use separate sessions for separate threads of work — one
for an incident, another for a routine check.

## Selecting credentials

Before the agent can touch a cloud account, you select which **credentials** the
session may use — for example a specific AWS account or GCP project. This is the
core of least-privilege in Nuphos:

* The agent only ever sees the credentials you select for that session.
* You can only select accounts you're allowed to use (the account's
  [member allow-list](/security/access-control)).
* Credentials are short-lived and minted on demand — see
  [the IAM trust model](/security/iam-trust).

<Tip>
  Give a session only the accounts it needs. If you're investigating staging,
  don't hand it production.
</Tip>

## Mentioning resources

You can reference Nuphos resources directly in your prompt with **mentions**,
instead of describing them in prose. Mention a cluster, a deployment, or an
account and the agent resolves it precisely — no ambiguity about which "api"
service you meant.

## How the agent works

As it runs, the agent moves through visible stages — building context, preparing
tools, thinking, and executing — and shows each tool call and result inline. It
has real tools at its disposal: a shell, cloud and Kubernetes APIs, web lookups,
and reusable [skills](/agent/skills).

When the agent wants to **change** something, it doesn't just do it — it proposes
a [plan](/agent/plans) for you to approve.

## History & long sessions

Conversations are persisted, so you can leave and come back. Long sessions are
**compacted** automatically — older turns are summarized while recent context is
kept verbatim — so a session can run for a long time without losing the plot.

## Beyond the desktop

The agent isn't limited to the desktop app. You can talk to it from
[Slack or iMessage](/collaboration/chat-agent), and run it automatically with
[triggers](/agent/triggers).