Lifecycle stages
These terms describe lifecycle stages for runtime versions:- Stable version: A released runtime version that supports production use. Each stable version includes a retirement date when support ends. The most current stable version is marked (Latest) — use it for new agents, agents in development, or when migrating from older versions.
- Legacy version: An older stable version that is supported but has been replaced by a newer stable version. Migrate to the latest version to access new features and improvements.
-
Preview version: A pre-release version that provides early access to some features. Do not use these versions in production systems.
- Private preview version: A preview version currently limited to specific projects selected by Cognite. Availability expands as Cognite validates the version.
- Retired version: A runtime version that is no longer supported and has reached its retirement date. Atlas AI automatically upgrades agents using a retired version to the latest stable version if you do not migrate them before the retirement date.
Agent runtime version overview
This table lists all available runtime versions with their current status and support timeline.Stable version features
This table lists the features available in each stable version.Preview version features
This table lists the features available in preview versions.Feature overview
These features vary between runtime versions.Query tool (Public preview)
Runtime versions 1.2.0 and 1.3.0 support the Query tool, which lets agents query Data Modeling directly with explicit operations for view discovery and instance reads. It can list available views, inspect view schemas, and run list, search, aggregate, and retrieve calls on instances with configurable data model and instance-space scoping.Subagents (Preview)
Runtime version 1.3.0 supports subagents, which allow an Atlas AI agent to delegate tasks to other configured agents. See About Atlas AI agent subagents for when to use subagents and their current limits.To-do list tracking
Runtime version 2.0.0 supports to-do list tracking, so an agent breaks a multi-step task into a list and tracks progress against each item. See About Atlas AI agent sandbox and context management for details.Sandbox
Runtime version 2.0.0 gives agents a sandboxed environment to execute commands and run scripts, with read access to your CDF project. See About Atlas AI agent sandbox and context management for details.Automatic context compression
Runtime version 2.0.0 automatically compresses long conversations to stay within the language model’s context limit. See About Atlas AI agent sandbox and context management for details.Tool result offloading
Runtime version 2.0.0 automatically stores large tool results outside the conversation and retrieves them again when needed. See About Atlas AI agent sandbox and context management for details.Query memory
Query memory is a caching capability that allows agents to store results from Query tools and pass the results as input to other tools in the same conversation. Some tools require Query memory to function, such as the Examine data semantically tool.Examine data semantically
This feature enables you to add the Examine data semantically tool to your Atlas AI agent.Select a runtime version
New agents automatically use the latest stable runtime version. You can select a different version in the agent builder user interface (UI) or through the Cognite API.- Agent builder UI
- Cognite API
The agent builder displays the runtime version in the agent configuration. Change the version from the dropdown.
Upgrade to a new version
Complete these steps before you upgrade your agent to a new runtime version:- Review the feature differences between your current version and the target version.
- Test your agent thoroughly in a non-production environment.
- Verify that any custom tools or integrations work correctly with the new version.
Atlas AI automatically upgrades agents to the latest stable version when their runtime version reaches the retirement date.