An engineer has used an AI coding agent to take a synthesizer from a blank folder to a working analog circuit board, and the hardware largely behaved the way the simulations predicted. The project ran about 19 hours from the first prompt to the order file, with the agent handling the circuit design, simulation, layout and documentation.

The author started with a simple question: could AI design real analog circuits in hardware? The release of GPT-6 Astra and Claude Opus 5.5 in September 2026 convinced him the tools had improved enough to test it. OpenAI had even shown a PCB designed by Astra in its announcement. He used Claude Code with Opus 5.5 in VSCode, working from an empty project and a fresh git repository.

He pointed the agent to the JLCPCB free parts list and asked it to propose what could be built from those parts. After the agent suggested several conventional options, the author picked an idea he found more interesting: a microcontroller-controlled analog synthesizer with touch keys on the board. The only other input was the microcontroller, a CH32V00x.

The agent proposed a subtractive synthesizer built from three main blocks: a Juno-style digitally controlled oscillator, a Moog-style transistor ladder filter and a voltage-controlled amplifier. The author set up the toolchain on Windows with Windows Subsystem for Linux for command-line tools. That included KiCad 10, ngspice for circuit simulation, pdflatex for documentation, the RISC-V toolchain and Python. Nearly everything the agent produced was a Python script.

The agent built its own component library from models it found on the web. It also corrected parts of its first design after reviewing the simulation results. Because a full simulation of the whole instrument was impractical at circuit level, the agent wrote a behavioral model in Python and calibrated it against the ngspice results. The audio samples the author published were generated with that model.

The author said he had little experience with analog synthesizers and was surprised by some of the choices the agent made. The parts selection was limited, so the design leaned on components that were common in the 1970s. He reviewed the design closely but did not change it. He noted that the oscillator relies on a resettable integrator, and that the agent chose a low-leakage op-amp and a high-voltage rated capacitor to reduce nonlinearity.

Layout took the most human intervention

The author said the layout stage needed the most help. The first version bunched most of the analog components together. The layout improved after he asked for a floorplan that followed the functional blocks. The agent wrote Python scripts for part placement and used Freerouting for routing.

Two-layer routing never fully completed, so the design moved to four layers. The author then pointed out how to shift circuit blocks to open routing channels, and after several rounds the board routed cleanly and passed design rule checks. He said a larger board than the 10 by 10 centimeter area he chose might have made the job easier.

The logs showed the agent spent most of its time on Freerouting runs that needed no interaction. Those runs, not the design decisions, accounted for most of the 19 hours.

Bring-up and the one hardware fix

The populated boards arrived about two weeks after the order. The author ran manual tests before installing the microcontroller, then connected the board to his lab PC through the programmer and an oscilloscope on the audio output. Claude Code controlled the lab PC over SSH, and the rest of the bring-up was automated through an audio loopback into the microcontroller.

The author said the match between SPICE simulations and the measured board was excellent. He said that surprised him, since the board was AI-designed and mismatches between simulated and real second-source components are common. The loopback also allowed automatic calibration of trim and control parameters.

Only one hardware fix was needed: removing a solder bridge on the microcontroller that the author had made by hand. The touch keys were initially too sensitive and needed recalibration. The author found them hard to use because of their small size, so he chose to control the synthesizer through microcontroller programs instead.

He used the oscilloscope to record the sound examples on his project page. Those tracks were generated automatically from sheet music and other sources, with multiple voices recorded separately and then combined.

In his conclusion, the author said generative AI can design PCBs with complex mixed-signal circuits. He said that will likely force changes to electronics design workflows, much as it is already doing in software. He added that he hopes it does not make electronics less interesting.

This article was written with the assistance of AI.
News Factory APP - agentic news to boost your SEO & AEO.