Source-linked AI summary
From Prompt to Prototype: Towards a Frontier LLM Driven RF Engineering Workflow
Markus Heinrichs, Oscar Moschner, Simon Tewes, Volker Wienstroer, Aydin Sezgin, Rainer Kronberger
TL;DR
RF hardware design lacks a demonstrated general workflow in which a frontier LLM autonomously drives professional tools from a specification through a complete product. This paper scripts CST, ADS, and KiCad to design and verify an active GNSS L1 antenna, achieving the reported antenna and prefilter results while shifting the engineer toward specification, trade-off, and review gates. The workflow produced a manufacturing-ready board design in simulation, but fabrication and measurement remain pending.
Problem
The paper addresses whether agentic coding environments can extend from software tasks to complete professional RF hardware design.
Method
A frontier LLM used scripting interfaces to operate CST, ADS, and KiCad across antenna, LNA, prefilter, PCB-layout, verification, and fabrication-data stages.
Results
A complete active GNSS L1 antenna was designed end-to-end in two hours, with all specifications met in simulation and reported antenna and prefilter performance.
Takeaways & Limitations
Agentic coding environments can make scripted RF hardware design automatable, reviewable, and reproducible while engineers specify, decide, and verify.
Takeaways & Limitations
All performance figures are simulations because the prototype is pending fabrication and measurement.
Abstract
from arXiv · showhide
Agentic coding environments give a frontier large language model (LLM) direct access to a workstation's terminal, file system, and software. This work demonstrates they extend to professional RF hardware design: an active GNSS L1-band antenna - a circularly polarized patch, surface acoustic wave (SAW) prefilter, and two-stage low-noise amplifier (LNA) on one printed circuit board (PCB) - was designed, optimized, and made manufacturing-ready. The LLM agent autonomously operated CST Studio Suite, Keysight ADS, and KiCad via scripting interfaces. Engineer input was limited to the specification, trade-off decisions, and design reviews. Workflow, results, and the RF engineer's evolving role are discussed.
I. INTRODUCTION
Agentic coding environments let frontier LLMs plan, execute, and verify tasks through workstation tools, and this paper applies that model to complete RF hardware design. Starting from one specification, the agent produced a manufacturing-ready active GNSS antenna.
- I. INTRODUCTION: Agentic coding environments give frontier LLMs direct access to terminals, files, and installed programs for iterative task execution.The agent writes code, runs it, inspects results, and iterates toward a specified goal.
- I. INTRODUCTION: RF design can be reformulated as a coding task because professional tools expose file formats, command-line executables, and Python APIs.This avoids dependence on custom-built tools or vendor-specific AI support.
- I. INTRODUCTION: The paper demonstrates the approach on a complete product design rather than an isolated RF subtask.The demonstration begins from a single specification.
- I. INTRODUCTION: The agent selected materials, designed the antenna and LNA, laid out the PCB, and generated fabrication data for an active GNSS L1-band antenna.The design combines a circularly polarized patch, SAW prefilter, and two-stage LNA on one PCB.
II. AGENTIC DESIGN WORKFLOW
The workflow fixes specifications and completion gates up front, then uses scripted RF tools and accumulated context to support autonomous iteration. The design state remains reproducible through scripts and parameter files.
- II. AGENTIC DESIGN WORKFLOW: The agent operates CST Studio Suite, ADS, and KiCad exclusively through scripting interfaces.These include CST Python libraries, ADS netlists executed by hpeesofsim, and KiCad’s pcbnew module and kicad-cli.
- II. AGENTIC DESIGN WORKFLOW: 2–3 s per circuit simulation and 3–5 min per fine-mesh full-wave simulation allow hundreds of simulate–evaluate–modify iterations.
- II. AGENTIC DESIGN WORKFLOW: Context files provide prior operating knowledge of CST and ADS so each new session need not relearn simulator automation.The files summarize the simulators’ Python interfaces and are maintained from previous sessions.
- II. AGENTIC DESIGN WORKFLOW: The prompt fixes PCB, powering, component-availability, antenna, and LNA requirements before development begins.The board is specified as a single two-layer PCB with a maximum outline of 120 mm × 120 mm and 5–12 V phantom powering.
- II. AGENTIC DESIGN WORKFLOW: Completion requires simulated specification compliance, clean ERC and DRC, visual layout inspection, and complete fabrication data.These criteria define when the agent’s autonomous loop may terminate.
- II. AGENTIC DESIGN WORKFLOW: Scripts and parameter files regenerate the schematic, board, checks, renders, and fabrication data with one command.
III. DESIGN CASE: ACTIVE GNSS L1 ANTENNA
The design case integrates the antenna, matching network, SAW prefilter, two-stage LNA, and SMA connector, while using each design step to expose the agent’s decisions and verification.
- III. DESIGN CASE: ACTIVE GNSS L1 ANTENNA: The final design integrates an RHCP corner-truncated patch, quarter-wave transformer, SAW prefilter, two-stage LNA, and edge-mounted SMA connector.The section presents the individual design steps to show how the agent reasons, decides, and verifies.
A. Substrate Selection and Antenna
The agent used simulation evidence to choose the substrate and debug circular-polarization optimization. The final antenna achieves the reported axial-ratio and realized-gain targets in simulation.
- A. Substrate Selection and Antenna: The design specification calls for a single two-layer PCB with a maximum 120 x 120 mm outline.
- A. Substrate Selection and Antenna: 28% radiation efficiency on FR4 led the agent to reject it, while Rogers RO4350B produced 78% efficiency.The comparison addressed dielectric-loss limits relative to the required 3 dBic realized gain.
- A. Substrate Selection and Antenna: The agent selected a corner-truncated square patch as the standard single-feed circular-polarization solution.
- A. Substrate Selection and Antenna: 2.46 dB axial ratio and 5.15 dBic realized gain at boresight were achieved after joint optimization of patch length and corner truncation.The correction addressed sequential one-parameter search and coarse frequency sampling.
B. Low-Noise Amplifier
The agent designed the prescribed BFP840FESD-based two-stage LNA and improved it through datasheet review, human review, and netlist comparison.
- Device choice: The LNA used the engineer-prescribed Infineon BFP840FESD transistor, narrowing the agent’s solution space.The specification deliberately mandated the component based on prior engineering experience, availability, qualification, or company standards.
- Biasing: The agent corrected an initially invalid bias point after detecting the ESD-protected variant’s VCEO = 2.25 V limit.The redesign followed unsolicited datasheet review during the project.
- Schematic review: Human review prompted a conventional signal-flow schematic with drawn wires, replacing a functionally complete but difficult-to-read grid arrangement.The agent rebuilt the schematic geometrically and compared old and new netlists as graphs rather than trusting the revised artifact.
- Verification: Fig. 4 presents the rearranged ADS schematic and simulated S-parameters for the final LNA.The caption identifies panel (a) as the schematic detail view and panel (b) as simulated S-parameters.
C. SAW Prefilter and System Performance
The agent incorporated an L1 SAW prefilter as a mid-project requirement, selecting a distributor-available filter by trading insertion loss for stronger out-of-band rejection while rechecking LNA compliance.
- Filter selection: The agent selected the TDK B4300 over a lower-loss alternative because 0.15 dB of additional insertion loss yielded 4–7 dB more rejection in threatened cellular and WiFi bands.The shortlisted GPS SAW filters were available at both mandated distributors.
- System verification: After adding the prefilter, the agent confirmed that the LNA still complied with all specifications.The simulated results and compliance summary appear together in Table 1.
- Requirement change: The prefilter requirement was handled as a design change request during the project.This demonstrates adaptation to changing requirements within the ongoing workflow.
D. PCB Integration and Fabrication Data
The scripted KiCad project preserved correspondence between simulation, schematic, layout, and fabrication outputs, while human visual review corrected physical implementation idealizations before release.
- PCB consistency: The generated schematic and layout were netlist-identical, with an automatic check verifying all 75 nodes and 18 nets against the simulated design of record.The project was generated from the scripted design description.
- Layout verification: RF copper polygons replicated the simulated line models, connecting the manufactured board to the design that was verified.Human visual review caught remaining deviations, including an incorrectly modeled second-stage collector junction.
- Design refinement: Human-requested refinements squared line endings, added a 45° feed-line miter, extended RF-line solder-mask openings, and added the SAW prefilter.Fig. 5 contrasts the first autonomous layout with the final board after review-driven refinement.
- Fabrication release: ERC and DRC passed without manufacturer-rule violations, all components were in stock at both mandated distributors, and complete assembly-format manufacturing data was delivered.These outcomes satisfied the prompt’s completion criteria.
IV. IMPACT ON THE RF ENGINEERING WORKFLOW
The workflow shifted RF engineering from direct tool operation toward specification, trade-off arbitration, and verification while one agent coordinated multiple professional design domains. The authors report reproducible, code-based outputs, but the evidence remains simulation-only and depends on documented interfaces and expert oversight.
- Workflow impact: The agent executed antenna design, circuit design, layout, and procurement-related work in approximately two hours without the engineer directly operating CST, ADS, or KiCad.The continuous context kept the domains mutually consistent.
- Autonomous engineering: The agent handled engineering steps requiring judgment on professional tools rather than only routine automation.This extends the demonstrated scope beyond simplified surrogate workflows.
- Engineer role: The engineer’s role concentrated on precise testable specifications, trade-off decisions, and verification through review gates backed by machine-checkable criteria.Expert oversight remained necessary because human reviews originated as-built fidelity corrections.
- Reproducibility: The design existed as version-controlled code that regenerated deterministically, supporting reproducibility and transferability.This was described as a side effect of the code-based workflow.
- Limitations: The results were limited to simulation because the prototype was pending fabrication and measurement, and the approach presumed documented scripting interfaces plus expert plausibility checks.The authors also report that reliable multi-hour sessions required frontier models released since early to mid-2026, in their experience.
V. CONCLUSION
The paper reports an end-to-end, simulation-compliant active GNSS L1 antenna design completed by a frontier LLM in two hours. It frames agentic coding environments as automatable, reviewable, and reproducible hardware-design workflows while retaining engineers for specification, decisions, and verification.
- Two hours was sufficient for a frontier LLM to design the complete active GNSS L1 antenna end-to-end.The design included the circularly polarized patch, SAW prefilter, two-stage LNA, and full fabrication data.
- All specifications are met in simulation.
- Agentic coding environments make hardware design automatable, reviewable, and reproducible when engineering tools expose scripting interfaces.
- The engineer shifts from operating tools to specifying, making trade-off decisions, and verifying through measurements.Measurement-based verification remains beyond the scope of an LLM agent.