An LLM Agent Drove Cadence and HFSS to Design a Sub-THz Transmitter

TL;DR

• Researchers Tristan Liang and Aravind Nagulu introduced TACS-RF, a framework where an LLM coding agent (Claude Opus 4.8) drives standard EDA tools to design RF and mmWave hardware.
• The test case was a 240 to 251 GHz transmitter in a 90 nm SiGe BiCMOS process.
• The paper reports +5.1 to +6.3 dBm output across an 11 GHz tuning range, 361 mW of DC power, and phase noise of -80 to -88 dBc/Hz at 1 MHz offset.
• The agent worked semi-autonomously. A human was still part of the loop.

An Agent at the Controls of Industry EDA Tools

The paper, submitted to arXiv on October 22, 2026, describes TACS-RF: Tool-driven Agentic Circuit Synthesis of RF and mmWave modules. The idea is simple to state. Instead of asking a language model to guess at circuit values, let it operate the same software a human designer would.

That software is serious. The agent, built on Claude Opus 4.8, drives Cadence Virtuoso and Spectre, Calibre, Cadence EMX, and Ansys HFSS. Those cover schematic work, simulation, layout verification and electromagnetic modeling.

The target was a sub-THz transmitter covering 240 to 251 GHz. The agent semi-autonomously designed sub-blocks including a 60 GHz push-push Colpitts VCO, multistage cascode power drivers, a push-push frequency doubler, and a differential bowtie antenna. By driving the tools directly, the framework automates sub-block synthesis, simulation, and matching networks.

The reported numbers: +5.1 to +6.3 dBm across an 11 GHz varactor tuning range, 361 mW DC power, and phase noise between -80 and -88 dBc/Hz at 1 MHz offset. The authors frame the work as “taking a step toward a ‘text-to-ASIC’ future for designing RF/mmWave hardware.”

Why This Isn’t Just Another Coding-Agent Demo

Most coding-agent stories end in a pull request. This one ends in a transmitter design. That’s a different kind of stakes, and I think it deserves a harder look than the usual benchmark victory lap.

Software agents work in a forgiving loop: write, run, read the error, fix. Here the loop runs through field solvers and circuit simulators, where a bad guess costs real simulation time. Can an LLM hold a design together across that many tools? The paper’s answer, at least for one design, is yes.

Think of a watchmaker who hires an apprentice with a dozen hands. The apprentice can work every bench tool at once, but the master still decides what the watch should do and checks the movement before it leaves the shop. That’s the shape of “semi-autonomous” as the paper describes it, and the authors say they cover the resulting division of labor between human and agent.

Here’s what the material I have doesn’t tell me. It doesn’t say how much of the work the human did, how long the agent took, or how the result compares with a design from an experienced team. Those gaps matter. A single transmitter in a single process, built by a single model, is a proof of concept and not a trend line.

But the choice of target is the interesting part. Sub-THz design sits at the awkward end of the spectrum, where manual simulation, EM modeling and tuning pile up. I’d argue that’s exactly where an agent that can grind through iterations has the most to offer. Or the most to get wrong.

My read: the headline isn’t that an AI “designed a chip.” It’s that an LLM operated a professional toolchain well enough to produce a coherent multi-block design. Smaller claim. More believable.

Why RF Design Is Slow in the First Place

RF and sub-THz integrated circuit design typically demands intensive manual simulation, electromagnetic modeling, and tuning across complex EDA toolchains. Nothing about that is quick. Each block interacts with the next, and the matching networks between them have to be coaxed into place.

That’s the work TACS-RF targets. The agent handles sub-block synthesis, simulation and matching networks, the repetitive loops that eat designer hours. The result is a chain running from a 60 GHz oscillator through a frequency doubler and power drivers to an antenna.

It also stretches what agentic coding means. The paper represents an expansion beyond software engineering into physical silicon hardware design and automated EDA tool orchestration. The code the agent writes is a means to drive tools, not the product itself.

What to Track After This Paper

First, look for detail on the human-agent split. The authors say they describe it, and that’s where the real answer to “how autonomous?” lives. If humans stepped in at critical decisions, the productivity story changes.

Second, watch for follow-ups on other circuit types or processes. One 90 nm SiGe BiCMOS transmitter shows the method can work. A second, different design would show it can travel.

Third, see whether the results get tied to physical hardware measurements or head-to-head comparisons with human-designed circuits. The material here doesn’t cover either, and I’d want both before calling this a workflow change.

Source: arXiv paper on TACS-RF