Comparison of isolated server and client-side runtimes for AI agents.

agent-runtimes.md

Agent Runtimes & Sandboxing

When developing agents, especially AI- or LLM-based, it's important to consider isolation, performance, and execution location (server vs client). We discussed the following concepts:

• Server-hosted runtimes: Node, Bun, Deno
• Server-side sandbox: Docker, gVisor, MicroVMs
• WebAssembly (WASM): server- or client-side
• WebContainers: full Linux-like environment in the browser
• Serverless/Edge runtimes: Cloudflare Workers, Deno Deploy, Vercel Edge Functions

---

1. Server-hosted Lightweight JS Runtimes

Node / Bun / Deno

• Runs directly on the server
• Fast startup, full API access
• No isolation: all agents share the same host
• Ideal for local experiments or trusted agents

Links

• Node.js
• Bun
• Deno

---

2. Server-side Sandbox

Docker / gVisor / Firecracker

• Strong isolation per agent
• Allows different Node versions and environments
• Higher resource usage
• Startup and management more complex than pure JS runtimes

Links

• Docker
• gVisor
• Firecracker

---

3. WebAssembly (WASM)

Server-side (Wasmtime / Wasmer)

• Lightweight sandbox, secure, isolated
• Fast, portable, can run multiple agents in parallel
• No Docker needed

Client-side

• Runs in the browser or WebContainers
• Fully sandboxed, offloads computation to user device
• Limited to browser APIs and device resources

Links

• WebAssembly
• Wasmtime
• Wasmer

---

4. WebContainers

• Full Linux-like environment in the browser
• Node/Bun/Deno runtime available
• CLI + GUI possible
• Agents run isolated, fully sandboxed
• Computation runs on the user device — great for demos, education, and benchmarking
• Free for small projects (StackBlitz)
• Commercial license needed for profit-oriented projects

Links

• StackBlitz WebContainers
• WebContainer API Docs

---

5. Edge / Serverless Runtimes

• Cloudflare Workers, Deno Deploy, Vercel Edge Functions
• Fast startup, isolated execution
• No server management required
• Limits on execution time and API access

Links

• Cloudflare Workers + WASM
• Deno Deploy
• Vercel Edge Functions

---

6. Summary / Recommendation

Approach	Isolation	Speed	Location	Use Case
Node / Bun / Deno	None	Very fast	Server	Local experiments, trusted agents
Node + Process Sandbox	Medium	Fast	Server	Multiple agents, moderate isolation
Docker / gVisor	Strong	Medium	Server	Untrusted agents, multiple environments
WASM	Strong	Very fast	Server or Client	Sandbox, lightweight agents, portability
WebContainers	Strong	Very fast	Client (Browser)	Sandbox + GUI + Terminal, education, benchmarking
Edge / Serverless	Strong	Very fast	Cloud	API calls, lightweight agents, scalable

Takeaways:

• Trusted, fast experiments → Node / Bun / Deno directly on server
• Secure, isolated agents → WASM on server or WebContainer in browser
• Client offload / demos / benchmarking → WebContainer or WASM in browser
• Production, untrusted agents → Docker / gVisor

---

7. Example: Node + WebContainer Setup

• Node web server runs in WebContainer
• Terminal + GUI available
• Agents can run isolated in the container
• Example:

// index.js
import express from 'express';
import { exec } from 'child_process';

const app = express();
const PORT = 3000;

app.get('/', (req, res) => res.send('Hello from WebContainer!'));

app.get('/run-agent', (req, res) => {
  exec('node agent.js', (err, stdout, stderr) => {
    if (err) return res.send(err.message);
    res.send(`<pre>${stdout}</pre>`);
  });
});

app.listen(PORT, () => console.log(`Server running on port ${PORT}`));