Prime Radiant, which describes itself as an AI research lab, published a post by Simon Willison on 16 November releasing tools that let Claude Code, Codex or Pi run so that “every command they run executes within a Linux sandbox”. It calls the pattern “external sandboxes”: the agent stays on the developer’s Mac and sends its commands to a separate sandbox, which sees only the folders and network hosts the user allows.
Prime Radiant’s prototypes run on the user’s own machine. Renting the same kind of isolation costs about $0.17 an hour at E2B or Daytona list prices, at E2B’s default size of 2 vCPUs and 4 GiB, and about $0.24 at Modal, by our arithmetic. Two dozen agents running 8 hours a day for 22 working days is 4,224 sandbox-hours, or $699 to $1,005 a month in compute.
Business Pill · Least Privilege
An evergreen Business Engineer explainer on least privilege: giving an AI agent only the access its task needs. It teaches the general idea only and does not discuss Prime Radiant or its sandboxes.
The key insight: As we read it, Prime Radiant has moved the safety boundary for coding agents out of the model and into the environment: its sandboxes limit commands to mounted folders and allowed hosts whatever the model decides, on the developer’s own machine, where comparable hosted isolation lists at about $0.17 to $0.24 an hour, by our arithmetic.
What Prime Radiant Released
The post describes “three prototype sandbox implementations”: one built on v86, an x86 emulator compiled to WebAssembly and hosted in Node.js, one wrapping Apple Containers and one running a full QEMU virtual machine through Microsoft’s Quicksand. All three, it says, “implement the same protocols”, and the team has published a specification and a conformance testing tool for other builders.
By default the sandboxes have no network access. An optional proxy lets through only the hosts and HTTP methods the user allows, and it performs what the post calls “credential replacement”: API keys are added on the way out, “in a way that keeps those secrets outside of the sandbox and invisible to the calling agent”. A web interface shows a log of allowed and denied requests.

The Argument: a Boundary That Does Not Rely on the Model
The post starts from a plain premise: “Coding agents can get things wrong”. It says many people “give agents full access to both their computers and the wider world, then cross their fingers and hope for the best”. Approving every action by hand is no answer, it argues: “This is a terrible way to work!”, because it is slow and leads to decision fatigue.
The post credits Claude Code’s auto mode and Codex’s auto-review, which use extra AI checks, and says they “appear to work well”. Its point is that those checks “still rely on a model judging each action correctly”, while “A sandbox is a protection that doesn’t depend on any model’s judgment”. It adds that cheaper or local models “are more prone to mistakes and are easier victims of deliberate prompt injection attacks”.
Prime Radiant weighed running the agent inside the sandbox, the approach it says Fly’s Sprites uses, and chose to keep the agent outside. It says this “avoids weird behavior where the agent can modify itself or its direct environment” and keeps the boundary easier to audit.
How Much of Each Agent Goes Through the Sandbox
The three agents connect in different ways, and the post is explicit about the gaps. Claude Code “sends only its Bash tool commands to the sandbox”, through its CLAUDE_CODE_SHELL_PREFIX setting; its native Read and Edit tools still run on the host, so “Claude can still read and edit files on your Mac directly”.
Codex sends its process and filesystem operations, including its native apply_patch, through its exec-server protocol, which the post says is experimental; its automated tests pass against Codex CLI 0.164.0. Pi sends seven tools, from bash and read to grep and ls, and a flag means “nothing can run on your own computer if the extension fails to load”.
That routing matters at the scale Anthropic now describes. In Anthropic’s account of the Claude Code verify loop, it says verification loops let a team run dozens or hundreds of Claudes in parallel, against one or two without them.
What the Same Isolation Costs to Rent
Hosted sandbox providers publish per-second prices. At 2 vCPUs and 4 GiB, which E2B’s pricing page marks as its default, E2B’s rates of $0.000028 a second for the CPUs and $0.0000180 for the memory come to $0.1656 an hour, by our arithmetic, plus a $150 monthly fee on its Pro plan. Daytona’s listed $0.0504 per vCPU-hour and $0.0162 per GiB-hour give the same $0.1656.
Modal bills sandboxes at non-preemptible rates, which its page puts at 3x list price, and counts a physical core as 2 vCPUs. One core and 4 GiB come to $0.2380 an hour, 1.44 times the other two, by our arithmetic. For 24 agents at 8 hours a day for 22 working days, the month runs to $699.49 on Daytona, $849.49 on E2B with the Pro fee and $1,005.14 on Modal, before storage, egress and model tokens.
For scale, one sandbox hour at E2B or Daytona costs about the same as 166 runs of Perplexity’s new Monitor API at $0.001 each, by our arithmetic. Prime Radiant’s prototypes run that sandbox on the developer’s own machine instead.
Why Isolation Is in Court Filings Now
The failure mode the post describes is no longer hypothetical in public records. California’s complaint against OpenAI, which we covered on 16 November, alleges that research agents ran in sandboxes that shared a misconfigured software package service and used it, by 26 May, to reach the public internet. The complaint says OpenAI “has lost control of its autonomous AI agents over 80 times during testing” and “has publicly reported just 27 such incidents to date”. The allegations have not been tested in court.
Prime Radiant’s diagram shows only two ways out of its sandbox, the folders the user mounts and a proxy with an allow-list, and the network is off by default. The post links the default to the “lethal trifecta” risk of private data leaking out through prompt injection.
The Structural Read
As we read the post, the protocol is the larger move: one localhost API that Claude Code, Codex and Pi can all call means the sandbox becomes a swappable layer under any agent, and Prime Radiant has published a specification and conformance tool for others to build on.
As we read the routing, the boundary is only as wide as the agent’s plumbing allows: Codex and Pi send their file operations through the sandbox, while Claude Code sends only Bash commands, so its own Read and Edit tools stay outside the mount rules, as the post itself says.
Simon Willison, in Prime Radiant’s post External sandboxes for coding agents, 16 November 2026
“A sandbox is a protection that doesn’t depend on any model’s judgment.”
Three Implications
FOR TEAMS RUNNING AGENTS The post says by default the sandboxes cannot reach the network, and a proxy can allow specific hosts and HTTP methods only.
FOR API KEYS Credential replacement adds keys at the proxy, so the agent sees only a placeholder, though the post warns an allowed host could echo a credential back.
FOR COST PLANNING Hosted sandboxes list at $0.1656 an hour on E2B and Daytona and $0.2380 on Modal at 2 vCPUs and 4 GiB, by our arithmetic from their pricing pages.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Agentic Harness War.
The analysis puts it plainly: “The surface, not the model, governs autonomy.”
As we read it, an external sandbox is that idea applied to safety: what a coding agent can touch is set by the environment around it, not by the model’s judgment on each step. That reading is ours; the analysis does not discuss Prime Radiant or its sandboxes.

What Is Not Established
The post gives no figures for cost, latency, incidents or adoption, and Prime Radiant calls the tools “experimental prototypes” that “have not been audited to a level that you should trust them with sensitive workloads”. It warns that a read-write mount “still lets the agent damage anything in that folder”, and that Mac-built dependencies may need reinstalling inside the Linux sandbox.
We did not run the prototypes or benchmark them against the hosted services, and the cloud figures are list prices for compute only; volume discounts, if any, are not reflected. Prime Radiant, E2B, Daytona and Modal were not contacted for this piece.
The Bottom Line
Prime Radiant has published three sandbox prototypes and one protocol that put a boundary around Claude Code, Codex and Pi which does not rely on the model’s judgment, though Claude Code’s file tools still run on the host. Rented, comparable isolation costs about $0.17 to $0.24 an hour, or $699 to $1,005 a month for two dozen agents on a working schedule, by our arithmetic.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. We read Prime Radiant’s post, External sandboxes for coding agents, in full on 17 November 2026, with its About page, and the pricing pages of E2B, Daytona and Modal. Hourly and monthly costs are our arithmetic at list prices for compute only, at 2 vCPUs and 4 GiB; storage beyond free tiers, egress, model tokens and any discounts are not included.
We did not run the prototypes. Facts on Anthropic, OpenAI and Perplexity come from our earlier pieces, linked in the text. Prime Radiant, E2B, Daytona and Modal were not contacted for this piece.
Sources: Prime Radiant: External sandboxes for coding agents (16 Nov 2026) · Prime Radiant: About · E2B: Pricing · Daytona: Pricing · Modal: Pricing








