---
title: 'In the News: November 15, 2026 (Extra 3)'
description: 'A hardware engineer had Claude Code design, simulate, lay out and bring up an analog synthesizer board in 19 hours from first prompt to order, working spec first.'
canonical_url: 'https://darkfactory.dev/news/2026-11-15-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-11-15-extra-3.md'
collection: news
date_published: '2026-11-15T23:05:00-05:00'
date_modified: '2026-11-15T23:05:00-05:00'
---

# In the News: November 15, 2026 (Extra 3)


A hardware blogger gave Claude Code a parts list and a microcontroller and got back a working analog synthesizer board, simulated before it was built and calibrated by the agent after it arrived. His write-up is mostly about method: generate each layer of the design from a complete specification, not by editing as you go. The second item is Alibaba's Qwen4-Max release, which matters here for its agent cost figures and an open-source plugin for long-running tasks.

## 1. An agent designs and brings up an analog synthesizer board from a parts list

**[Agentic AI Designs an Analog Synthesizer](https://cpldcpu.github.io/2026/11/14/touchsynth/)** · Tim, hardware and microcontroller blogger (cpldcpu) · Tim's Blog, November 14, 2026

Tim started from a blank folder and a fresh git repo, running Claude Code with Opus 5.5. He pointed it at JLCPCB's free parts list and named the microcontroller (CH32V00x); the agent proposed a subtractive synthesizer with a Juno-style oscillator, a Moog-style transistor ladder filter and a voltage-controlled amplifier. It then ran seven ngspice test benches, built a Python behavioral model calibrated against the SPICE results, generated KiCad schematics, and wrote the firmware and ordering files. By his logs, the design took 19 hours from first prompt to order, most of it unattended autorouting runs. PCB layout was where he intervened most. He had to ask for a floorplan that followed the functional blocks, two-layer routing never completed, and the four-layer board cleared only after he pointed out where to move blocks to make room for routing.

His working rule is the useful part. "My observation is that it is much easier for the models to generate the next abstraction layer from a complete specification in one go, instead of doing it step by step," he writes, and he has the agent write a design report in parallel. He also had the agent add an audio feedback path into the microcontroller's ADC so bring-up could be automated. When the boards arrived, Claude Code drove the lab PC over SSH and calibrated the board through that loopback and an oscilloscope. He reports an excellent match between simulation and hardware, with one hardware fix needed: a solder bridge he had made by hand. The design files, the AI-written report and audio samples are published.

**Why it matters:** This is a first-person account of agent work outside software that follows the same discipline as a software factory: full spec per abstraction layer, a simulator as the check before anything physical, and a test path designed in so the agent can verify its own output. The 19 hours and the hardware match are his figures; the published repo is there for anyone who wants to check the design.

## 2. Qwen4-Max ships with lower agent costs and a plugin for staged delivery

**[Qwen4-Max: Your New Teammate Has Arrived](https://qwen.ai/blog?id=qwen4-max)** · Qwen Team, Alibaba · qwen.ai, November 2026

Alibaba's Qwen team released Qwen4-Max, a sparse mixture-of-experts model with a 2.4-trillion-parameter backbone, 71B parameters active per token and a one-million-token context window. The figures that bear on agent workloads are about tokens and price. Against Qwen3.8-Max, the team reports token usage on Terminal-Bench 4.0 down about 48% and 49% at its Xhigh and High effort settings, lower API prices with "particularly large reductions in cached-input pricing," and average cost per task on its in-house Work360 benchmark falling from $4.43 to $1.67. The post also announces an open-source Qwen-Live-Delivery-Plugin for long-running agent tasks: the agent ships a usable first version early, then revises it in successive releases while the user can step in and redirect.

On generated-code security, the team reports that the share of function-level outputs passing both functionality and security checks on Last Secure Code Bench rose from 22.0% to 31.8%. The model was built alongside harness and product teams including Qoder, and is available on QwenCloud by API. Every benchmark and cost figure above is the vendor's own, and Work360 is a benchmark Alibaba built. On Hacker News the post had 165 points and 39 comments about two hours after submission.

**Why it matters:** Alibaba has priced Qwen4-Max below Qwen3.8-Max, most sharply on cached input, and reports it needs about half the tokens on terminal agent tasks, so teams running agent loops on Qwen have reason to re-price them. The Live Delivery idea, staged releases with the human able to intervene throughout, is a harness pattern worth trying regardless of which model runs it.
