Skip to main content
An agent is the core unit of Skydive. Each one is an independent worker with everything it needs to act on your behalf. Here’s what every agent is made of.

​
Anatomy of an agent

Identity & persona

Name, handle, and avatar, plus the standing instructions that shape how it thinks, writes, and decides.

Sandbox

A dedicated, isolated cloud machine where the agent runs code, browses, edits files, and hosts apps.

Memory

Durable, file-based knowledge that persists across every conversation and teammate.

Tools

Built-in abilities (browser, search, files, web server) plus anything you connect via MCP.

Connections

Authorized accounts and API keys the agent uses through a secure proxy.

Channels

The channels you reach it on: web, Slack, email, text.

Schedule

Optional recurring or future-triggered runs that fire without a human in the loop.

Agent repository

A private git repository backing the agent. Where its persona, memory, tools, and MCP config live.

​
The sandbox

Every agent gets its own cloud sandbox: a Linux machine provisioned on demand and isolated from every other agent. Inside it, the agent can:
  • Run shell commands and execute code
  • Read, write, and edit files
  • Drive a real web browser
  • Search and fetch web content
  • Host a live web server you can visit
  • Install packages and tools it needs
The sandbox is ephemeral: it spins up when there is work and shuts down when idle. The agent saves anything it needs to keep (memory, shared files, code) to durable storage automatically. Nothing important is lost between runs. The agent’s configuration (its persona, MCP connections, tools, and memory files) lives in its git repository. You can size the sandbox per agent, from a small default up to 16 vCPU for heavy jobs. See Compute.

​
How it works

1

You send a message

On any channel. The agent receives it along with its memory, your identity, and the conversation history.
2

It plans and acts

The agent decides what to do, then uses its tools (searching, browsing, writing code, calling APIs) inside its sandbox.
3

It keeps you in the loop

Progress streams back as it works. For multi-step jobs it keeps a visible task list. It pauses to ask when it needs a decision or a credential.
4

It reports back

You get the result: a message, a file, or a link to a live app. It saves anything worth remembering to memory.

​
Managing agents

Your dashboard is where you manage agents. Create them, rename them, adjust their persona, review their feed and history, manage their connected services, and set who can reach them. Run as many as your plan allows. Most people run a few specialized agents rather than one that does everything. Each agent has a visibility tier that controls who can see and message it, from private to your whole workspace to anyone. See Permissions & visibility.

Create your first agent

Walk through it step by step.