The evolved Agents SDK gives the model a powerful harness with instructions, tools, approvals, tracing, handoffs, resume bookkeeping, and the kind of model behavior used in Codex-style agents.
The sandbox provides compute: files, commands, packages, artifacts, and isolation. These two layers can run on the same machine or separately for better isolation, durability, and security. That separation also helps keep credentials and orchestration outside the environment where model-generated code runs.
A sandbox agent is still an Agent. The API stays the same, but the agent now gets a workspace you can customize with a Manifest, along with sandbox-native capabilities such as:
-
Filesystem tools
-
Shell access
-
Skills
-
Memory
-
Compaction
This makes it possible to bring data to the sandbox instead of forcing everything into context.
The Manifest adds a portable way to define the agent workspace, so the same workflow can move more easily across providers and environments. It can:
-
Stage local files
-
Clone repositories
-
Create output directories
-
Mount external storage such as S3, GCS, Azure Blob Storage, and Cloudflare R2
-
Work with Git repositories
Permissions let you define sandbox users, attach file permissions to Manifest entries, and run model-facing tools as a specific user. Under the hood, this maps to standard Unix filesystem permissions. For example, the model can have read-only access to dataroom/ while keeping write access to output/.
For teams deploying agents at scale, durability matters. A long-running agent should not lose its working state because a container expired, crashed, or a workflow was paused. With built-in snapshotting and rehydration, the Agents SDK can restore a session in a fresh sandbox and continue from its last saved state instead of starting over.
Jump straight into the code.