Swap coding agents and sandbox providers without changing code. Preserves full interactive capabilities (approval flows, streaming).

Key Takeaways

• Swap coding agents and sandbox providers without changing code.
• Preserves full interactive capabilities (approval flows, streaming).

What Changed

AgentBox is a new SDK that abstracts the runtime for coding agents. Instead of wrapping claude --print (non-interactive mode), it launches each agent as a server process inside a sandbox and communicates over WebSocket or HTTP. This preserves approval flows, tool-use control, and streaming events.

Key abstraction: One API for any agent + any sandbox provider.

import { Agent, Sandbox } from "agentbox-sdk";

const sandbox = new Sandbox("local-docker", {
  workingDir: "/workspace",
  image: process.env.IMAGE_ID!,
  env: { ANTHROPIC_API_KEY: process.env.ANTHROPIC_API_KEY! },
});

const agent = new Agent("claude-code", {
  sandbox,
  cwd: "/workspace",
  approvalMode: "auto",
});

const result = await agent.run({
  model: "sonnet",
  input: "Create a hello world Express server in /workspace/server.ts",
});

await sandbox.delete();

What It Means For You

If you're building multi-agent workflows or need to run Claude Code in a CI/CD pipeline, this matters. Most existing solutions call agents in non-interactive mode (claude --print), which strips away approval flows and tool-use control. AgentBox preserves the full interactive session.

Supported agents:

• claude-code
• opencode
• codex

Supported sandboxes:

• local-docker
• e2b
• modal
• daytona
• vercel

Swap either — your app code stays the same. This is particularly useful for:

• Running untrusted agent code in isolated environments
• Parallelizing agent runs across multiple sandboxes
• Testing different agents on the same task without refactoring

Try It Now

1. Install: npm install agentbox-sdk (requires Node >= 20)

2. Build a sandbox image:
   

   npx agentbox image build --provider local-docker --preset browser-agent

This prints an image reference. Set it as IMAGE_ID.

1. Stream events in real-time:

   const run = agent.stream({
     model: "sonnet",
     input: "Write a fizzbuzz in Python",
   });

   for await (const event of run) {
     if (event.type === "text.delta") {
       process.stdout.write(event.delta);
     }
   }

   const result = await run.finished;

1. Key methods on sandbox: run(), runAsync(), gitClone(), openPort(), getPreviewLink(), snapshot(), stop(), delete()

gentic.news Analysis

AgentBox arrives at a time when Claude Code usage is surging — it appeared in 58 articles this week alone (total: 634 across our coverage). The trend toward running agents in sandboxed environments aligns with the recent CVE-2026-35022 security disclosure for Claude Code, which highlighted the risks of running agents without isolation.

This SDK directly addresses a pain point we've seen in our coverage: developers want to use Claude Code in CI/CD but need proper sandboxing. Previously, they had to choose between non-interactive mode (losing approval flows) or custom scripting. AgentBox provides a standardized abstraction similar to what the Vercel AI SDK did for LLM calls — but for agent + runtime.

The ability to swap between Claude Code, Codex, and OpenCode without changing code is particularly valuable as the agent ecosystem fragments. With Claude Opus 4.6 scoring 94.1% on ThermoQA and Codex 5.3 competing on SWE-Bench, having a provider-agnostic runtime lets you benchmark agents on your actual tasks.

What you should do differently: If you're currently running Claude Code with claude --print in CI, migrate to AgentBox for sandboxed, interactive sessions. If you're building multi-agent architectures, use AgentBox as your runtime abstraction layer — it'll save you from rewriting integration code when you switch sandbox providers or agents.

Originally published on gentic.news