# Agent Sandbox — Run Agents Safely on Kubernetes

> Agent Sandbox provides Kubernetes-first guardrails for agent workloads: resource limits, isolation, and repeatable environments so failures stay contained.

## Install

Save as a script file and run:

## Quick Use

1. Install / run:
   ```bash
   git clone https://github.com/kubernetes-sigs/agent-sandbox && cd agent-sandbox
   ```
2. Start / smoke test:
   ```bash
   sed -n '1,260p' README.md
   ```
3. Verify:
   - Deploy the smallest example, then confirm: (1) CPU/memory limits are enforced, (2) network/credentials are scoped, (3) logs capture every run.

## Intro

Agent Sandbox provides Kubernetes-first guardrails for agent workloads: resource limits, isolation, and repeatable environments so failures stay contained.

- **Best for:** Teams running untrusted agent code who need isolation, quotas, and an auditable execution surface
- **Works with:** Kubernetes clusters + CI; pair with policy-as-code and least-privilege service accounts
- **Setup time:** 30 minutes

## Practical Notes

- Setup time ~30 minutes (apply manifests + run one sandboxed job)
- Three hard checks: limits enforced, egress scoped, run logs retained
- GitHub stars + forks (verified): see Source & Thanks

Agents get dangerous when they can run arbitrary code with long-lived credentials. A Kubernetes sandbox lets you bound blast radius: small quotas, short-lived identities, and auditable logs. Use it to make ‘agent execution’ an infrastructure primitive rather than an ad-hoc local script.

### FAQ

**Q: Do I need Kubernetes for agents?**
A: Not always—but it’s a strong default if you must isolate untrusted execution.

**Q: What’s the first guardrail to add?**
A: Resource limits + restricted service accounts; then add egress controls.

**Q: How do I make runs reproducible?**
A: Pin images/versions and treat manifests as code reviewed like any PR.

## Source & Thanks

> Source: https://github.com/kubernetes-sigs/agent-sandbox
> License: Apache-2.0
> GitHub stars: 2,125 · forks: 253

---

<!-- ZH -->

## 快速使用

1. 安装 / 运行：
   ```bash
   git clone https://github.com/kubernetes-sigs/agent-sandbox && cd agent-sandbox
   ```
2. 启动 / 冒烟测试：
   ```bash
   sed -n '1,260p' README.md
   ```
3. 验证：
   - 部署最小示例后，验收三点：① CPU/内存限额生效；② 网络/凭证被收紧；③ 日志能完整记录每次运行。

## 简介

Agent Sandbox 提供 Kubernetes 优先的运行模式：通过资源限额、隔离与可复现环境，把 agent 工作负载关进“可控沙箱”，并用配额、策略与审计把失败影响面压到最小，长期可持续。

- **适合谁：** 要跑不完全可信的 agent 代码，并需要隔离/配额/可审计执行面的团队
- **可搭配：** Kubernetes 集群 + CI；适合配合策略即代码与最小权限账号
- **准备时间：** 30 分钟

## 实战建议

- 上手约 30 分钟（apply 清单 + 跑一个沙箱任务）
- 三项硬性验收：限额生效、外联受控、运行日志留存
- GitHub stars + forks（已核验）：见「来源与感谢」

当 agent 能拿着长期凭证执行任意代码时，风险会指数级上升。Kubernetes 沙箱能把爆炸半径变小：小配额、短期身份、可审计日志。把“agent 执行”做成基础设施能力，而不是临时脚本。

### FAQ

**跑 agent 一定要上 Kubernetes 吗？**
A: 不一定，但当你需要隔离不可信执行时，它是很强的默认方案。

**第一道护栏应该加什么？**
A: 资源限额 + 限制 service account；然后再加外联控制。

**如何保证可复现？**
A: 固定镜像/版本，把清单当代码走 PR 审核。

## 来源与感谢

> Source: https://github.com/kubernetes-sigs/agent-sandbox
> License: Apache-2.0
> GitHub stars: 2,125 · forks: 253


---
Source: https://tokrepo.com/en/workflows/agent-sandbox-run-agents-safely-on-kubernetes
Author: AI Open Source