Nov 22, 2025 · 49m · latent-space
⚡️ Building the AI Hardware Engineer with Matthias Wagner, Co-founder of Flux
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Matthias Wagner, co-founder and CEO of Flux AI, discusses building an AI-powered hardware engineer within a browser-based CAD platform, demonstrating how autonomous agents synthesize circuit schematics and connect directly to component supply chains.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 4.3 Guest teaching 3.7 Guest disagreement 0.1 The hosts pushing back 0.2
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 43:31 Dismissing the necessity of mass manufacturing for low cost
Matthias forcefully counters the traditional assumption that mass production is required to make electronics inexpensive, predicting that autonomous design tools will dismantle the traditional OEM model.
Hardest push from the hosts ▶ 36:07 Skepticism on the prompt management startup categoryThe host pushes back on the viability of standalone prompt management businesses, citing recent startup shutdowns and noting that top engineering teams build customized internal tooling instead.
Biggest teaching moment ▶ 10:15 Explaining supply chain bottlenecks in PCB designMatthias educates the host on how a single out-of-stock component among hundreds can delay an entire hardware project by 200 days, demonstrating why live distributor integration is vital.
The host holds their own ▶ 36:40 Sharing Cognition's proprietary prompt playground architectureThe host demonstrates deep technical domain knowledge by detailing how Cognition independently solved prompt versioning and tracing workflows for Devin two years prior.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Origins of Flux and Rethinking Hardware Tooling | 2 | 3 | 0 | 0 | The host opens with an exploratory question about the concept of an AI hardware engineer. Matthias shares his background from Siemens to Meta and how the stagnation in hardware tooling sparked Flux. | |
| From RAG Chat to Autonomous Hardware Agents | 3 | 4 | 0 | 0 | The host asks whether GPT-3 immediately unlocked Flux or if deeper engineering was needed. Matthias details the progression from early RAG chatbots to reliable tool-calling and autonomous execution plans. | |
| Exploring Flux UI and Supply Chain Integration | 4 | 4 | 0 | 0 | Matthias demos the Flux UI and live supply chain lookup. The host asks technical clarifying questions regarding manufacturing export formats, API integrations, and volume pricing dynamics. | |
| Agent Architecture, Search Pipelines, and Dynamic Memory | 4 | 4 | 0 | 0 | The host inquires about MCP versus custom function calling and vector search implementations. Matthias details their hybrid search architecture and dynamic user-level memory system. | |
| Live Demo: Designing a Voice Assistant Device | 5 | 3 | 0 | 0 | The host shares personal experience trying to build embedded hardware projects and proposes UX enhancements like inline wiki tooltips, referencing his work on Windsurf code maps. | |
| Developer Tooling Synergy, Computer Use, and WebGL | 7 | 3 | 0 | 1 | The host reveals he works directly on Devin and discusses incorporating computer use for canvas manipulation. Matthias shares direct technical pain points with browser GPU virtualization and drag events. | |
| Autonomous Component Placement and Open Hardware Libraries | 3 | 4 | 0 | 0 | The live agent searches and selects hardware components autonomously while Matthias explains Flux's open-source, user-generated GitHub-like library model and board fabrication pipelines in Shenzhen. | |
| AI Engineering: Evals, Prompt Management, and LangSmith | 8 | 4 | 0 | 1 | The host engages in a high-level peer discussion on prompt engineering, noting how Cognition internally solved prompt CMS and tagging challenges, while analyzing why prompt management startups struggle. | |
| Commercial Growth, PLG Strategy, and Future Manufacturing | 3 | 5 | 1 | 0 | Matthias details commercial milestones (7,000 paying users, 26x growth) and articulates an ambitious long-term thesis on AI replacing conventional OEM mass-manufacturing models. | |
| Live Demo Review, Recruitment, and Final Thoughts | 4 | 3 | 0 | 0 | The host and guest review the completed schematic output from the live demo, concluding with discussions on WebGL performance challenges and recruitment goals. |
Statements from this episode (17)
Wagner: Hardware design tooling has barely improved since the mid-nineties
“And I had this realization that the tooling to make hardware had not improved in my lifetime. And that was so odd because the tooling to make software is like unrecognizable if you compare it to like mid nineties or something like that, right?”
Building a custom IDE would cost AI coding startups $100M
“If like Cursor or Devon would have to build their own VS code, that'd cost a hundred million dollars, you know? Call me back in five years.”
Wagner claims Flux shipped embedded AI chat before GPT-4 released
“So I think I'm going to claim here, I think we were the first engineering tool or design tool that had an AI chat in it. We shipped that I think a month or two months before GPT-IV became publicly available.”
Wagner: Reliable LLM tool calling in late 2023 unlocked functional agents
“And I would say like that tool calling, I think really only became reliable about a year ago, like last October, November, when the models that shipped were suddenly really good at calling tools. And I think that then changed everything, right? And you see tha…”
Flux shipped an autonomous agent for hardware board design in October
“And so that then led us to Really like heavily invest in the agenda capability in Flux, right? And going from like a, hey, look, I can answer questions and perform single tasks to like, no, I can take a product brief, break that down into an actual execution p…”
One out-of-stock component can delay an entire hardware project 200 days
“And when you have a design with say, 500 components, and one of them is not on stock or there's a lead time of 200 days, then the whole project has a lead time now of 200 days, right?”
Flux's AI agent finds electronic component replacements in under a minute
“This is going to take here, 30, 60 seconds, hopefully. And it's like, this would have taken us hours to do manually.”
Flux boasts a user-generated library of over one million electronic parts
“So we have our own library for all these parts, over a million. That's all user generated.”
Wagner: Task-specific knowledge injection effectively reduces LLM hallucinations
“The way to, like, get them to behave the way we want them to behave is just with context engineering and injecting the right knowledge for the right kind of tasks they're doing. And that's incredibly effective, right? That reduces hallucinations, that makes th…”
Wagner says Devin's 'Ask' function is its most underrated feature
“Now we use Devon here religiously, and I think one of the most underrated features of Devon is the actual ask feature.”
Wagner: AI agent output quality heavily depends on upfront requirement specifications
“I think a lot of success depends how well you flesh out what you wanna do, right? Like if you just use this and ask it to like start here executing a plan, it's gonna be high variability in what you're gonna get back, right? But the more we go through here and…”
Wagner admits Flux's knowledge-base architecture was inspired by Cognition's Devin
“We love Devin here. Yeah, I mean, I know a lot here is inspired by Devon, like the knowledge-based thing, right, is inspired by Devon, right?”
Wagner: Fine-tuned open-source models can beat frontier models on quality
“I think you're now with, you know, we'll be seeing the trend going also with the coding agents, you know, fine-tuning open source models and like way faster as a result and more accurate. You can actually beat frontier models on quality that way.”
Wagner: Prompt management startups fail because the problem isn't well-understood
“I also think the reason why these startups haven't done well so far is I don't think we understand the problem well enough yet. To abstract a generalizable solution. Right. And I think that's where everybody is stuck.”
Flux's autonomous hardware engineering agent runs for 25 minutes on average
“So yeah, the average session is I think, 25 minutes.”