Field notes from European Microwave Week 2026, London.
One of the most interesting screens at the Keysight booth at EuMW 2026 had no trace on it at all. It showed a prompt. Under the "Pioneer 6G Research" banner, an AI coding agent was driving a real RF bench: a signal generator, a signal analyser and a small connectorised amplifier, all through natural-language instructions. The prompt is worth reading closely, because the most important engineering content in it is not about AI. It is a power budget.
The setup
The wall display showed a document titled "ZX60-24A-S+ bench prompts". It describes two independent workflows, each written as a self-contained prompt. The one on screen was "Harmonics (single tone, through 5th order)". It defines the bench as follows:
- Signal analyser: a Keysight SA6210A, with a maximum input of +30 dBm.
- Signal source: a Keysight N5186A, with a maximum CW frequency of 8.5 GHz.
- Signal path: N5186A RF Output 1 → Mini-Circuits ZX60-24A-S+ → SA6210A RF Input 2.
- DUT: the ZX60-24A-S+ low-noise amplifier. The prompt gives its gain as 24 dB, its input damage level as +20 dBm and its output-referred P1dB as about +18 dBm.
- Output plumbing: about −3 dB of cable and adapter loss at 5 GHz.
The agent's tool calls were visible in the left pane. A tool labelled "Smarti [kdk_search]" was being queried for "Harmonics measurement setup SCPI" and "CW output setup: preset, make sure ARB…". In other words, the agent looked up instrument knowledge before issuing commands. The signal generator's front panel read "System is under remote control". The model selector in the agent window showed a Claude Sonnet model.
The power budget is the guardrail
The last paragraph of the prompt gives the agent its authority:
"Drive is −10 dBm at the DUT input, 30 dB below its damage rating, and the analyzer sees about +11 dBm, 19 dB below its +30 dBm maximum, so there is no credible damage path."
The arithmetic checks out: −10 dBm + 24 dB gain − 3 dB loss = +11 dBm. On the strength of that margin, the prompt tells the agent not to pause for confirmation and pre-authorises preset, RF output enable and any level up to the stated drive.
This is the real lesson of the demo. An agent that can enable an RF output is only as safe as the numbers it has been given. Here the engineer did the damage analysis once, wrote it down explicitly, and then allowed the agent to run without interruption. That is a sound pattern. Three items deserve to be in every such prompt:
- The absolute maximum of every port in the path, as given in the prompt for the DUT input and the analyser input.
- The gain and loss between them, so the agent and the reviewer can recompute the levels.
- A hard ceiling on drive, stated as a number, not as "be careful".
Safe is not the same as accurate
The budget guarantees that nothing gets damaged. It does not guarantee that the harmonics measured are the amplifier's own. Two points from classic harmonic measurement practice apply directly:
1. The amplifier is driven close to compression. With +24 dB gain at −10 dBm input, the output is about +14 dBm, roughly 4 dB below the stated output P1dB of about +18 dBm. That is a perfectly legitimate operating point for a harmonic test, but it should be a deliberate choice. Harmonic levels change quickly near compression, so a single drive level gives one data point, not a characterisation. A drive sweep, which an agent can run cheaply, is more informative.
2. The analyser can generate its own harmonics. About +11 dBm reaches the analyser input. Unless the input attenuation is set high enough, the analyser's first mixer sees a large signal and produces harmonics of its own, which add to or mask the DUT's. The standard checks are:
- Change the input attenuation by 10 dB and see whether the harmonic levels relative to the carrier change. If they do, at least part of what you see comes from the analyser.
- Know the source's own harmonics. The signal generator's harmonics pass through the amplifier together with the fundamental. A low-pass filter at the source output, or a measurement of the source alone, separates them.
- Keep the harmonics within the analyser's frequency range. Through the 5th order of a 5 GHz tone, that means 25 GHz.
None of these checks is beyond an agent. They need to be written into the workflow, in the same explicit style as the power budget. "Verify that the measured harmonics do not change when the input attenuation is increased by 10 dB" is exactly the kind of instruction that turns an automated measurement into a trustworthy one.
What this demo says about AI on the bench
The interesting part was not that an AI can type SCPI. It was the division of labour. The engineer supplied the physics: device limits, path loss, safe drive. The agent supplied the execution: instrument lookup, command sequencing and data collection. The instrument vendor supplied a knowledge tool the agent could query. That is a sensible split, and it puts responsibility where it belongs. When something goes wrong, the first place to look is the prompt, because that is where the engineering judgement lives.
Takeaway
An AI agent driving an RF bench is a real capability now, not a mock-up. The demo also showed what makes it trustworthy: a written power budget with numbers that anyone can recompute. The next step is to write the measurement-validity checks with the same discipline, so the agent is not only safe but also right.
Notes: Observations and interpretations are my own, based on what was running at the booth during the show. Bench definition and DUT figures are quoted or paraphrased from the prompt shown on the booth display. Instrument settings such as input attenuation were not legible in my photos and were not confirmed with the exhibitor.