---
title: 'In the News: November 16, 2026 (Extra 3)'
description: "Daniel Stenberg counts the curl fixes that agent-built Rust rewrites missed; a new tool keeps agents' credentials on the laptop, approved per request."
canonical_url: 'https://darkfactory.dev/news/2026-11-16-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-11-16-extra-3.md'
collection: news
date_published: '2026-11-16T23:05:00-05:00'
date_modified: '2026-11-16T23:05:00-05:00'
---

# In the News: November 16, 2026 (Extra 3)


While agents turned 185,000 lines of curl's C into 340,000 lines of Rust, the real curl project merged more than 300 commits, about 200 of them bug fixes, by Daniel Stenberg's count. Also in this edition: a developer who runs his coding agents on throwaway VMs has released a tool that keeps his credentials on his laptop and runs the commands that need them there, one approval at a time.

## 1. Curl's maintainer counts what the agent-written Rust rewrites already lack

**[Actual curl rewrites in rust](https://daniel.haxx.se/blog/2026/11/17/actual-curl-rewrites-in-rust/)** · Daniel Stenberg, curl maintainer · daniel.haxx.se, November 17, 2026

Stenberg points to two public Rust reimplementations of curl: [urlx](https://github.com/jonwiggins/urlx), which claimed to be a memory-safe, from-scratch reimplementation, and [a November rewrite](https://github.com/kov/curl) that keeps the name curl and, he notes, claims to have spent 8,600 USD. He counted the result: agents converted 185,000 lines of C into 340,000 lines of Rust in a little more than three weeks. Over the same weeks the curl project merged over 300 commits, a little over 200 of them bug fixes, plus fixes for a few vulnerabilities. None of those bug fixes are in the rewrites. He credits the speed to the material agents had to work from: curl is extremely well documented, hundreds of open source applications use its API, and thousands of test cases can check an implementation. "Writing code once was never the big deal with software and it still isn't," he writes. He also puts a number on the memory-safety argument: of the 81 vulnerabilities curl has confirmed so far in 2026, 14 (17.3%) would most likely have been avoided in a language such as Rust, so a Rust curl would still have had 67.

**Why it matters:** A mature project with full documentation and a large test suite is close to an ideal specification, which is why agents can reproduce it in weeks. Stenberg's figures price the part they do not reproduce: curl averages about 11 commits a day, and a fork has to absorb that stream or fall behind. Anyone costing an agent-driven rewrite should budget for that maintenance as well as the conversion.

## 2. Keeping credentials on the laptop while agents run on remote VMs

**[Running coding agents without handing them my credentials](https://blog.antonio.cx/running-agents-with-lend)** · Antonio, pico.sh · antonio's blog, November 13, 2026

Antonio runs Claude Code, Codex and longer-running agents on disposable VMs, and keeps his SSH keys, his 1Password vault and his logged-in `gh` and cloud CLIs on his laptop. He rejects the two usual fixes: copying tokens onto the VM, where every process can read them, and forwarding the SSH agent with `ssh -A`, which lets the VM use every key "without asking and without any record of what it signed." His tool, [lend](https://github.com/antoniomika/lend), sits in front of the existing SSH agent. A remote agent asks to run a command through `lend exec` or an MCP tool, `run_local_command`, that requires a reason. The laptop checks the request against a default-deny policy, asks for approval (Touch ID on macOS), runs the command locally and streams back stdout, stderr and the exit code. Policy rules can inject secrets such as `GH_TOKEN` from 1Password into an approved command without the VM ever seeing them, hide or limit SSH keys, cap signatures per session and grant short leases. A `PermissionRequest` hook sends Claude Code's ordinary permission prompts to the same dialog. He lists the limits himself: approved commands run unsandboxed as his user, approving a shell or interpreter amounts to approving anything, the caller details in the prompt come from the VM and can be faked, and the tool is early and runs on macOS and Linux only.

**Why it matters:** A default-deny allowlist with per-request approval, executed where the credential lives, keeps long-lived tokens out of the agent's environment entirely. Antonio's own caveat sets the design rule: approval only limits what runs if the approved command cannot run arbitrary input, so the allowlist has to name narrow commands and never a shell.
