Skip to content
Strategy Agent Docs

The agent runs code

5 November 2026

The agent now runs code instead of estimating:

  • When an answer needs arithmetic or a comparison, the agent writes Python or a shell command and runs it in a Sandbox. Each run shows as a card among the steps of the answer, with the code, its output and how long it took. If a run fails, the agent fixes its code and runs it again. See Let the agent run code.
  • A run ends after about five minutes, and Stop ends it at once. The Sandbox restarts after that and every 30 minutes, so the agent rebuilds what it needs.
  • Files you want to keep go into the folder of the Conversation on your Personal Drive and stay when you delete the Conversation. The card of a run lists the files it wrote, each with a Download. The agent reads your Personal Drive with its file tools, under /personal. Earlier versions of a file are kept for 365 days, and your Personal Drive is deleted when you are removed from the Tenant.
  • Code reaches the internet only if you hold the new Right Sandbox has internet access. Both preset Roles, Admin and Member, hold it. Admins see a warning wherever it is shown or granted, a note at the Tenant’s first sign-in, and the Right can be removed from both Roles. A change applies from the next run, and taking the Right away stops running Sandboxes within moments. See Internet access for code and Set up your Tenant.
  • The Trust Center describes how credentials are hidden in the Sandbox and why that is a safeguard, not a boundary, what the code can reach on your Personal Drive, and how we trace its network traffic. See Security and Privacy.