Why agents need sandboxes, not larger context windows

What this post added

This post introduces the concept of agent sandboxes as a solution to the limitations of LLM context windows for long-running agent tasks. It details how sandboxes provide agents with a virtual environment (file system, command line, code runtime) to act as short-term memory, enabling them to process larger datasets and perform complex analyses. The post explains how sandboxes integrate with Glean's Enterprise Graph and leverage a Python runtime for data processing. It also discusses lessons learned from previous memory management techniques, such as data classification and selector tools, and how these inform the sandbox design. Finally, it outlines the security measures implemented for sandboxes, including isolation and access controls.