◉ANTIGRAVITY LABJP
Articles/Agents & Manager
◈ Agents & Manager/2026-06-19Advanced

Your Antigravity Sandbox Isolates Multi-Agents Less Than You Think — Notes on Containing the Blast Radius

An Antigravity sandbox gives you the feeling of isolation, but isolation leaks through three real gaps: shared volumes, over-broad allowed domains, and approval fatigue. Field notes on plugging the leaks, containing the blast radius by design, and proving isolation holds with tests.

antigravity476multi-agent50sandboxsecurity23production71

✦ Premium Article

You feel isolated, but only the blast radius is growing

The moment I turned on Antigravity's sandbox, I relaxed. "Now, whatever an agent does, the damage stays inside the box." That relaxation turned out to be the most dangerous part.

I run several blogs as a personal developer, driven by agents working in parallel. At one point I wired up a setup where a content-generation agent handed artifacts to a neighboring agent through a shared directory. The sandbox was on. The filesystem was isolated. Or so I assumed. Then one agent overwrote a temporary file another had placed in that shared directory, and an unpushed draft was gone. The sandbox worked perfectly. I was the one who had drilled an escape hatch through isolation — by sharing a directory I treated as a free-for-all workspace.

A sandbox gives you the feeling of isolation. But isolation isn't finished the instant you enable it. The job is to separate what's actually protected from the places you punch holes yourself, and to keep the blast radius small by design. Here are the leaks I hit in practice and the code I use to close them.

Separate what the sandbox protects from what it doesn't

The first move is to drop your over-expectations. As of mid-2026 (the v2.1.x line), Antigravity's sandbox protects roughly this set by default.

Protected for youNot protected (you design it)
Direct writes to real project files (work happens on a copy)Cross-overwrites of data placed in shared volumes
Outbound traffic to domains you didn't allowHow "wide" an allowed domain is — allow one, and everything under it passes
Processes escaping the sandboxOperators clicking "approve all" in the approval dialog
Resource ceilings (process count, memory, runtime)Credentials shared and reused across agents

The runtime handles the left column. The trouble lives in the right column, and that is entirely operations and policy design. What my deleted draft taught me is that leaks only happen on the right. Let's close them in order.

✦

Thank you for reading this far.

Continue Reading

What follows includes implementation code, benchmarks, and practical content we hope you'll find useful. This site runs without ads — server and development costs are supported entirely by members like you. If it's been helpful, we'd be truly grateful for your support.

WHAT YOU'LL LEARN
✦Separate what the sandbox protects from what it doesn't, and plug the three places isolation actually leaks
✦A deny-by-default per-agent permission policy that keeps the blast radius minimal
✦Containment assertions that prove isolation holds before you ever launch the agents
Secure payment via Stripe · Cancel anytime
✦

Unlock This Article

Get full access to the rest of this article. Buy once, read anytime. This site is ad-free — your support goes directly toward keeping it running.

or
Unlock all articles with Membership →
Share

Thank You for Reading

Antigravity Lab is ad-free, supported entirely by members like you. We publish practical guides daily with implementation code, benchmarks, and production-ready patterns. If you've found it useful, we'd love to have you on board.

  • ✦Copy-paste ready implementation code
  • ✦New advanced guides published daily
  • ✦$5/mo or $10 for lifetime access
View Membership →

Related Articles

◈ Agents & Manager2026-04-22
Prompt Injection Defense in Antigravity — A Production Security Playbook for LLM Apps
Build a four-layer prompt injection defense inside Antigravity with working Python. Includes measured heuristic recall across eight obfuscations and the normalization bug that breaks detection for non-Latin scripts.
◈ Agents & Manager2026-06-01
Capping Parallel Agents With a Token Budget — Designing a Guard That Stops Runaway Cost
Running many agents in parallel quietly inflates your token bill. This is not about shrinking prompts — it is about a governance layer that meters spend in real time and cuts it off at a budget. Full design and TypeScript implementation, drawn from my own overnight jobs.
◈ Agents & Manager2026-05-27
Record & Replay for Antigravity Agents
How to deterministically replay a failed Antigravity Agent run offline, drawn from a month of running it across four production sites. Covers boundary recording, R2 + KV storage costs, PII masking, and a working TypeScript harness.