deepagents v0.4 ships pluggable sandbox support, smarter conversation history summarization, and Responses API as the default for OpenAI models.

Agents can now execute code in fully isolated environments through native integrations with Modal, Daytona, and Runloop. Configure your sandbox backend in a few lines and your agent gets filesystem tools (ls, read_file, write_file) plus an execute tool for running arbitrary shell commands — all without touching your host system, credentials, or local files.

There are two ways to integrate: run the agent inside the sandbox for tight environment coupling, or run it on your server and call the sandbox as a remote tool (our recommended approach — API keys stay outside, you pay only for execution time, and you can update agent logic without rebuilding images). Read our architecture deep dive for a full breakdown of the trade-offs.

Get started with the sandboxes documentation