Agent Sandbox
An agent sandbox is a limited place where an AI agent can run code or tools. It separates the task from the rest of the computer, but protection depends on its design.
harness | sandbox → reviewed artifacttemporary copy · no secretsA coding agent works in a temporary copy of a repository. Network access is off, secrets are absent, and time and disk are limited.trusted controlnetwork off · 60 sdiff + auditisolated files/process/network + resource limits + export gateOPSAnother example
A coding agent works in a temporary copy of a repository. Network access is off, secrets are absent, and time and disk are limited.
Main parts
- 01Isolated file space
- 02Process boundary
- 03Network policy
- 04Credential policy
- 05Resource and time limits
- 06Export gate and log
Use it when
Use it when an agent needs an isolated workspace, local tool state, or risky code execution. Limit files, network, credentials, processes, time, and exports.
Do not use it when
Do not call a normal folder a sandbox. Isolation does not replace permission checks, monitoring, or human review.
Name used in code
isolated files/process/network + resource limits + export gate Before you ship
Check the model and prompt versions, source data, versioned evaluation set, and measures. Verify links to evidence, tool permissions, privacy, common failures, decline-to-answer and fallback behavior, human review, monitoring, cost, speed, and rollback.
Outcome: Use or evaluate Agent Sandbox to make AI behavior measurable and tied to evidence. User context: A coding agent works in a temporary copy of a repository. Network access is off, secrets are absent, and time and disk are limited. AI method or concept: Agent Sandbox. Why it fits: Use it when an agent needs an isolated workspace, local tool state, or risky code execution. Limit files, network, credentials, processes, time, and exports. Do not use it when: Do not call a normal folder a sandbox. Isolation does not replace permission checks, monitoring, or human review. AI requirements: Define the input and source evidence. Set model and tool permissions. Use versioned evaluation data and measures. Define failure, decline-to-answer, privacy, speed, and cost limits. Operational safeguards: Keep a clear trace and hide sensitive log data. Show users a safe fallback. Mark steps that need human review. Define how to roll back the model or prompt. Acceptance criteria: Record baseline and target measures on a versioned evaluation set. Test edge cases and hostile inputs. Verify fallback, monitoring, permissions, and rollback. Evidence and limits (evidence boundary): It is official for that source. Other platforms or teams may use the term in another way. Unknowns to confirm: Target task, model and version, evaluation owner, source data, risk limit, tool permissions, and production fallback.
Check this request
Review the current use of Agent Sandbox. Definition: An agent sandbox is a limited place where an AI agent can run code or tools. It separates the task from the rest of the computer, but protection depends on its design. Release checks: Check the model and prompt versions, source data, versioned evaluation set, and measures. Verify links to evidence, tool permissions, privacy, common failures, decline-to-answer and fallback behavior, human review, monitoring, cost, speed, and rollback. Before changing code, report the evidence you found, gaps, severity, and the smallest safe fix.