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

# About Atlas AI agent sandbox and context management

> Understand the to-do list tracking, sandbox execution environment, automatic context compression, and tool result offloading capabilities available in newer Atlas AI agent runtime versions.

<Warning>
  These capabilities are available starting with [agent runtime version 2.0.0](/cdf/atlas_ai/references/atlas_ai_agent_runtime_versions) (Private preview). Availability is currently limited to select projects. Do not rely on preview features in production systems.
</Warning>

Longer, more complex agent tasks, such as analyzing data across thousands of rows or time series data points, can run into a few problems: the task needs to be broken down and tracked, some tools need to run code or commands in a safe environment, and long conversations can grow past the language model's context limit. Runtime version 2.0.0 addresses these with **to-do list tracking**, a **sandbox execution environment**, and **automatic context compression**, so agents can run longer, more complex tasks.

## To-do list tracking

For a multi-step task, the agent breaks the work into a to-do list and tracks progress against each item as it works. This produces more reliable results for tasks that span many steps, such as processing a large dataset in batches.

## Sandbox

The sandbox gives an agent an isolated, temporary environment to execute commands and run scripts that use the [Cognite Python Software Development Kit (SDK)](/dev/sdks/python/).

From the sandbox, an agent can read data from your CDF project but can't create, update, or delete it. Using the sandbox requires the `sessions:create` [capability](/cdf/access/guides/capabilities#atlas-ai).

Each command times out after 9.5 minutes. The sandbox stays available for about 12 hours after the last command, then clears automatically.

## Automatic context compression

When a conversation grows long, the agent compresses older turns while retaining the details needed to continue the task. This prevents long-running and multi-step tasks from failing when they approach the language model's context limit. No configuration is required.

## Tool result offloading

When a tool call returns a large result, such as a lengthy document or extensive query output, the agent stores it separately and keeps only a short preview in context. If the full result is needed later, the agent retrieves it automatically. No configuration is required.

## Related information

* [Agent runtime versions](/cdf/atlas_ai/references/atlas_ai_agent_runtime_versions)
* [About Atlas AI agent subagents](/cdf/atlas_ai/concepts/atlas_ai_agent_subagents)
* [About Atlas AI agent tools](/cdf/atlas_ai/concepts/atlas_ai_agent_tools)
* [Capabilities](/cdf/access/guides/capabilities#atlas-ai)