From Prompt to Prototype: Towards a Frontier LLM Driven RF Engineering Workflow

I. INTRODUCTION

• Agentic coding environments give frontier LLMs workstation terminal, file-system, and program access to plan, execute, and verify multi-step tasks, extending software-engineering iteration beyond code writing.Claude Code is cited as a widely used implementation.Source: §I · p. 1
• Coding agents can script professional RF software through documented file formats, command-line executables, and Python APIs, making design state regenerable without vendor AI integration.This recasts RF design as a coding task rather than GUI interaction.Source: §I · p. 1
• Earlier hardware AI approaches include RF-chip synthesis, digital-design agents spanning synthesis through tape-out, and proprietary EDA assistants, all requiring custom tools or vendor support.Source: §I · p. 1
• From one specification, Claude Fable 5 designed and prepared for manufacture a complete active GNSS L1 antenna integrating a circularly polarized patch, SAW-prefiltered two-stage LNA, and PCB.The agent used CST for the patch, ADS for the LNA and prefilter, and KiCad for the layout and fabrication data.Source: §I · p. 1, Figure 1 · p. 1

II. AGENTIC DESIGN WORKFLOW

• The agent used CST Python libraries, ADS netlists via hpeesofsim, and KiCad’s pcbnew and kicad-cli interfaces, enabling hundreds of iterations at 2–3 s per circuit simulation and 3–5 min per fine-mesh full-wave simulation.It ran on a standard engineering workstation with CST Studio Suite 2026, ADS 2024, and KiCad 9.Source: §II · pp. 1–2
• Because new LLM sessions lack prior memory, two short, iteratively maintained Markdown context documents supplied simulator-interface knowledge, while design methodology and component selection came from the model.The documents summarized the Python interfaces for CST and ADS.Source: §II · p. 2
• The prompt fixed the boundary conditions before development: a single two-layer PCB up to 120 mm × 120 mm, phantom power of 5–12 V over the RF port, and components from mandated distributors.Antenna and LNA functional requirements were specified separately in Section III.Source: §II · p. 2
• Completion required simulated compliance, a KiCad project with clean ERC and DRC consistent with the simulated design, visual layout inspection, and complete fabrication data; these gates also defined when the autonomous loop could terminate.Source: §II · p. 2
• The agent maintained design state in scripts and parameter files, allowing the schematic, board, design-rule checks, renders, and fabrication data to regenerate with one command.Source: §II · p. 2

III. DESIGN CASE: ACTIVE GNSS L1 ANTENNA

• The final design combines an RHCP corner-truncated patch, quarter-wave transformer, SAW prefilter, two-stage LNA, and edge-mounted SMA connector; the subsections use its design steps to show agent reasoning, decisions, and verification.The specification and simulated results are listed in Table 1.Source: §III · p. 2

A. Substrate Selection and Antenna

• On Rogers RO4350B, simulated radiation efficiency rose to 78%, compared with 28% on standard FR4, leading the agent to select the RF laminate.The FR4 result was too low for the required 3 dBic realized gain.Source: A. Substrate Selection and Antenna · p. 2
• After joint two-dimensional optimization, the final antenna achieved 2.46 dB axial ratio and 5.15 dBic boresight realized gain.The agent optimized patch length and corner truncation at the exact center frequency after identifying sequential one-parameter-at-a-time search and coarse frequency sampling as the problem. An auxiliary probe-fed model ruled out the microstrip feed as the cause.Source: A. Substrate Selection and Antenna · p. 2, Figure 3 · p. 3
• For circular polarization, the agent chose a corner-truncated square patch, the standard single-feed solution.The substrate choice was deliberately left open, so the agent generated simulation evidence rather than relying on designer experience.Source: A. Substrate Selection and Antenna · p. 2

B. Low-Noise Amplifier

• The specification mandated an Infineon BFP840FESD, narrowing the agent’s design choices based on the engineer’s prior experience.The passage notes that component availability, qualification, or company standards can also motivate such specification choices.Source: B. Low-Noise Amplifier · p. 3
• The agent redesigned the bias after noticing its initial choice violated the BFP840FESD’s VCEO limit of 2.25 V.It found the constraint while reviewing the transistor datasheet mid-design, without being asked.Source: B. Low-Noise Amplifier · p. 3
• After engineer feedback, the agent redrew the schematic in signal-flow order with wires and checked the revised netlist against the original as graphs.The first schematic was functionally complete but machine-oriented; persistent context files let the instruction become a standing convention for future projects. Fig. 4 presents the rearranged ADS schematic and final-LNA simulated S-parameters.Source: B. Low-Noise Amplifier · p. 3, Figure 4 · p. 3

C. SAW Prefilter and System Performance

• The agent chose TDK B4300, accepting 0.15 dB additional insertion loss for 4–7 dB more rejection in threatened cellular and WiFi bands.The agent confirmed that the LNA still complied with all specifications.Source: C. SAW Prefilter and System Performance · p. 3

D. PCB Integration and Fabrication Data

• All 75 nodes and 18 nets were automatically checked against the simulated design, and copper-polygon RF lines replicated its line models.Engineer visual review caught and corrected idealizations, including an amplifier-stage collector junction simulated as a crossing but routed as a tee; the simulation was updated to use microstrip tee MT_C2.Source: D. PCB Integration and Fabrication Data · pp. 3–4
• The final board passed ERC and DRC without violations, used in-stock components at both mandated distributors, and was delivered in the assembly service’s manufacturing-data format.Compared with the first autonomous layout, review squared the line endings, added a 45° feed-line miter and extended solder-mask openings around RF lines; the final board also included the SAW prefilter.Source: D. PCB Integration and Fabrication Data · p. 4, Figure 5 · p. 4

• In about two hours, one agent completed work spanning antenna and circuit design, layout, and procurement in a continuous context without the engineer operating CST, ADS, or KiCad.No familiarity with the tools’ user interfaces or scripting APIs was required, and the continuous context kept domains mutually consistent.Source: §IV · p. 4
• The agent performed judgment-intensive engineering on professional tools, while the engineer focused on precise, testable specifications, trade-offs, and verification.Expert review remained load-bearing: human design reviews prompted the as-built fidelity corrections, shifting oversight from continuous tool operation to review gates backed by machine-checkable criteria.Source: §IV · p. 4
• The design was version-controlled code that regenerates deterministically, making it reproducible and transferable compared with most interactive CAD sessions.This was a side effect of the workflow.Source: §IV · p. 4
• All performance figures were simulations pending prototype fabrication and measurement; the approach presumes documented scripting interfaces and an engineer able to judge intermediate results.The authors report that, in their experience, only frontier models released since early to mid-2026 reliably sustain multi-hour tool-driving sessions; they say comparable agentic coding environments from other trillion-parameter-class models should also be able to drive the workflow.Source: §IV · p. 4

V. CONCLUSION

• In two hours, a frontier LLM designed an active GNSS L1 antenna end-to-end with fabrication data, meeting all specifications in simulation.The design comprised a circularly polarized patch, SAW prefilter, and two-stage LNA.Source: §V · p. 4
• Where engineering tools expose scripting interfaces, hardware design can be automatable, reviewable, and reproducible without vendor-side AI integration.The paper argues that agentic coding environments therefore extend beyond software development.Source: §V · p. 4
• The engineer remains in the loop, shifting from tool operation to specifying, deciding, and verifying through measurements.Measurement-based verification is described as beyond the scope of an LLM agent.Source: §V · p. 4