<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/making-an-ai-agent-design-circuit-boards-that-actually-work-a-deterministic-check-after-every-llm-s-yglimxr42" -->

---
title: Making an AI agent design circuit boards that actually...
description: An AI agent for PCB design gets reliable results by checking every LLM step with deterministic tools rather than trusting the model's output. The agent, called...
canonical: https://daily.dev/posts/making-an-ai-agent-design-circuit-boards-that-actually-work-a-deterministic-check-after-every-llm-s-yglimxr42
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: Making an AI agent design circuit boards that actually work: a deterministic check after every LLM step | daily.dev
og:description: An AI agent for PCB design gets reliable results by checking every LLM step with deterministic tools rather than trusting the model's output. The agent, called...
og:url: https://daily.dev/posts/making-an-ai-agent-design-circuit-boards-that-actually-work-a-deterministic-check-after-every-llm-s-yglimxr42
og:image: https://api.daily.dev/og/posts/YGliMXr42.png
og:image:alt: Making an AI agent design circuit boards that actually work: a deterministic check after every LLM step
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Making an AI agent design circuit boards that actually work: a deterministic check after every LLM step

**[Build With GenAI](https://daily.dev/sources/buildwithgenai)** · [@stefano4](https://daily.dev/stefano4) · 2 min read · 19 upvotes · 1 comments

## Summary

An AI agent for PCB design gets reliable results by checking every LLM step with deterministic tools rather than trusting the model's output. The agent, called PCBBrew, has the model write a netlist first, which a checker validates for pin types, missing power rails and voltage mismatches. KiCad then generates the schematic, and its netlist export is diffed against the model's version to catch drift. Footprints come from real KiCad libraries instead of model guesses, firmware runs against WebAssembly-emulated peripherals wired to the same netlist, and analog circuits are checked in ngspice against values the model predicts beforehand. Placement and routing use classic algorithms, with DRC as the final check. The author found cheap models work fine when the checks are strong, and forcing a single round of clarifying questions up front shortened conversations significantly.

## Content

LLMs are surprisingly good at electronics *reasoning* and surprisingly bad at the details that make a board work: pin numbers, footprints, which GPIO can be an input at boot. For the last months I've been building an agent that goes from a written project description to a manufacturable PCB, and the main lesson was: **never trust a step the model can't have checked by something deterministic.**

What the loop looks like:

- **Netlist first.** The model writes a netlist with explicit "interfaces" between blocks (who supplies which rail, voltage ranges, buses). Before anything is drawn, a checker validates it: ERC on pin types, rails without a source, voltage mismatches.
- **The schematic is verified, not trusted.** KiCad generates the schematic and its own netlist export is diffed against the model's netlist. Any difference goes back to the model as an error.
- **Parts come from real libraries.** The model searches symbol/footprint libraries through tools, and the tool now returns the footprint KiCad itself associates with a symbol. That alone killed a whole class of "plausible but wrong package" mistakes.
- **Firmware runs against the circuit.** The firmware is compiled for the real MCU and executed in WebAssembly emulators wired to the same netlist: displays, I2C sensors, encoders, motors. If the code talks to the wrong pin, the emulated peripheral doesn't answer, and the model sees that in the log.
- **Analog goes through SPICE.** Power and audio parts are simulated with ngspice; the model has to state expected values and the simulation either confirms them or sends it back.
- **Placement and routing are classic algorithms**, not the LLM: the model chooses *what*, deterministic code decides *where*, and DRC has the final word.

Things that surprised me:
- Asking the model to declare expected measurements *before* simulating made its designs noticeably better.
- Models love to ask questions one at a time; forcing a single round of questions up front (and blocking a second round in code) made conversations much shorter.
- Cheap models are fine for most steps when the checks are strong; the checks matter more than the model.

The project is PCBBrew (https://pcbbrew.dev, private beta). Short video: https://youtu.be/71zhhrTaoFE

Curious how others structure verification in agent pipelines outside of code generation.

## Community discussion

Top comments from developers on daily.dev.

**@amananand755** · 1 upvotes

> 'The model chooses what, deterministic code decides where, and DRC has the final word' is a great rule for any agent pipeline. Diffing KiCad's netlist against the model's and running firmware against emulated peripherals are clever checks.

## Similar posts on daily.dev

- [Eye-Catching Consumer Tech at CES 2026](https://daily.dev/posts/eye-catching-consumer-tech-at-ces-2026-b5mojomhr) · IEEE Spectrum · 1 upvotes · 0 comments
- [Cadence Develops Agentic AI Platform for PCB Design](https://daily.dev/posts/cadence-develops-agentic-ai-platform-for-pcb-design-ki2cak297) · Embedded.com · 0 upvotes · 0 comments
- [Cadence opens the door to chips designed for AI by AI](https://daily.dev/posts/cadence-opens-the-door-to-chips-designed-for-ai-by-ai-k8vtousbc) · The Register · 0 upvotes · 0 comments
- [Large Tabular Models Excel Where LLMs Fail](https://daily.dev/posts/large-tabular-models-excel-where-llms-fail-m4afd4hqb) · IEEE Spectrum · 0 upvotes · 0 comments

---

Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#llm](https://daily.dev/tags/llm), [#hardware](https://daily.dev/tags/hardware)

[View this post on daily.dev](https://daily.dev/posts/making-an-ai-agent-design-circuit-boards-that-actually-work-a-deterministic-check-after-every-llm-s-yglimxr42)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"DiscussionForumPosting","mainEntityOfPage":"https://daily.dev/posts/making-an-ai-agent-design-circuit-boards-that-actually-work-a-deterministic-check-after-every-llm-s-yglimxr42","headline":"Making an AI agent design circuit boards that actually work: a deterministic check after every LLM step","text":"An AI agent for PCB design gets reliable results by checking every LLM step with deterministic tools rather than trusting the model's output. The agent, called PCBBrew, has the model write a netlist first, which a checker validates for pin types, missing power rails and voltage mismatches. KiCad then generates the schematic, and its netlist export is diffed against the model's version to catch drift. Footprints come from real KiCad libraries instead of model guesses, firmware runs against WebAssembly-emulated peripherals wired to the same netlist, and analog circuits are checked in ngspice against values the model predicts beforehand. Placement and routing use classic algorithms, with DRC as the final check. The author found cheap models work fine when the checks are strong, and forcing a single round of clarifying questions up front shortened conversations significantly.","url":"https://daily.dev/posts/making-an-ai-agent-design-circuit-boards-that-actually-work-a-deterministic-check-after-every-llm-s-yglimxr42","datePublished":"2026-11-18T17:21:48.129Z","dateModified":"2026-11-18T17:22:15.648Z","author":{"@type":"Person","name":"Stefano","url":"https://daily.dev/stefano4","image":"https://lh3.googleusercontent.com/a/ACg8ocLbsdalNAatSvPdo7sqQARTy3x89-oRNO3P2qptSoEqo0QDVWU=s96-c","interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"EndorseAction"},"userInteractionCount":70}},"image":"https://media.daily.dev/image/upload/s--UdZ4wsKD--/f_auto,pg_1/v1795022508/posts/YGliMXr42?_a=BAMAMiiw0","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":19},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"comment":[{"@type":"Comment","text":"‘The model chooses what, deterministic code decides where, and DRC has the final word’ is a great rule for any agent pipeline. Diffing KiCad’s netlist against the model’s and running firmware against emulated peripherals are clever checks.","datePublished":"2026-11-19T07:45:11.070Z","url":"https://daily.dev/posts/YGliMXr42#c-5UXj2uGMh","author":{"@type":"Person","name":"Aman Anand","url":"https://daily.dev/amananand755","image":"https://media.daily.dev/image/upload/s--IC7nsnTH--/f_auto,pg_1/v1794242586/avatars/avatar_HwTBgfjy0?_a=BAMAMiiw0"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1}}],"isPartOf":{"@type":"WebPage","url":"https://daily.dev/squads/buildwithgenai","name":"Build With GenAI"}}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Build With GenAI","item":"https://daily.dev/squads/buildwithgenai"},{"@type":"ListItem","position":3,"name":"Making an AI agent design circuit boards that actually work: a deterministic check after every LLM step"}]}
```

