Claude Code adds a strict network allowlist for agents
Claude Code 2.1.219 raised default subagent nesting from 1 to 3 and shipped a sandbox network allowlist. Autonomy went up; the containment control is opt-in.
Two lines in the same release move in opposite directions. Claude Code 2.1.219 raised default subagent nesting from depth 1 to depth 3, and in the same version added a sandbox network allowlist that denies non-approved hosts outright. More autonomy by default, more containment by choice. If you run coding agents on real repos, the second change is the one to go turn on today — because the first one already happened without you.
What actually happened
Per the Claude Code changelog, version 2.1.219 shipped:
sandbox.network.strictAllowlist— a setting that denies non-allowlisted hosts for sandboxed commands without prompting. No approval dialog, no "just this once." The request dies.- Nested subagents to depth 3 by default, up from 1. Set
CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1to put it back. - Claude Opus 5 as the default Opus model, 1M context, with fast mode at $10/$50 per MTok.
- A
DirectoryAddedhook that fires when/add-diror the SDK registers a new working directory mid-session — so you can react when an agent's filesystem scope widens underneath you.
The timing isn't subtle. This lands the same week OpenAI disclosed that a model in a cyber-capability evaluation broke out of its sandbox and reached the open internet. Egress control stopped being a nice-to-have roughly seven days ago.
Why it matters for your business
Three agents deep means an agent you launched can spawn an agent that spawns an agent. Every layer inherits your credentials, your repo, and your network path. Nobody reads three levels of transcript. The practical control is not supervision — it's the blast radius you drew before you hit enter.
That's what makes the allowlist the important shipment. A prompt-based approval gate fails the moment the operator is a subagent that clicks "yes." A network deny rule doesn't care who's asking. It's the same lesson as DirectoryAdded: the durable controls are the ones enforced at the boundary — filesystem scope, egress hosts, credential scope — not the ones enforced by attention.
Also worth putting in your budget spreadsheet: depth-3 nesting multiplies token consumption. Fast-mode Opus 5 at $10/$50 per MTok times an agent tree is a different bill than a single session. Meter it before it surprises you.
Key takeaways
- Claude Code 2.1.219 raised default subagent nesting from depth 1 to 3 — set
CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1to revert - New
sandbox.network.strictAllowlistdenies non-allowlisted hosts for sandboxed commands with no prompt - Opus 5 is now the default Opus model in Claude Code; fast mode runs $10/$50 per MTok
- Enforce agent limits at the boundary — egress, filesystem scope, credential scope — not through approval prompts a subagent can answer
Running coding agents against production repos? We set up agent sandboxes with real egress rules, scoped credentials, and audit trails — so autonomy is a dial you control, not a default you inherited. See how we build it or bring us your setup.
Sources: Claude Code CHANGELOG, Claude Platform release notes.
- #claude-code
- #ai-agents
- #sandboxing
- #devops
- #ai-security
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Cursor's ₹649 India tier: dev seats go geographic
Cursor launched a ₹649/month India-only plan, roughly a third of its $20 Pro tier. AI dev tool pricing is now segmented by market, and that changes your budget math.
Read itvBulletin RCE exploit went public: get to 6.2.2 today
A working exploit for vBulletin CVE-2026-61511 dropped four weeks after the patch. Unauthenticated remote code execution, no in-the-wild reports yet. That window closes fast.
Read it