2025-12-02 –, Walnascar
After a few months of using Cursor AI backed by claude-4-sonnet in agentic mode on Yocto / OpenEmbedded development, I want to share the genuine productivity transformations I've experienced, and the critical lessons about where it goes wrong.
I'll show you how AI helped me remote into target hardware via SSH, capture webcam images of running code, analyse the output, and iterate until our e-ink display worked correctly. I'll demonstrate automated power consumption monitoring feeding back into BSP optimisation workflows.
I'll also share the gotchas: when AI creates overly complex solutions, goes down rabbit holes, and happily does the nonsensical things if you ask it to.
The key insight? You need feedback loops, you need experience, and you need to stay in control,
I'm hoping to begin a conversation with the community about how we build best practice to leverage AI pair programming in future.
The Productivity Revolution
Here's what using "AI" - by which I mean Cursor backed by claude-4-sonnet in agentic mode - has genuinely transformed in my workflow:
Research and Problem-Solving
AI searches multiple sources, combines results, and presents options orders of magnitude faster than I can Google. I'm constantly learning new approaches to problems I thought I understood.
Tool Mastery
AI knows ways to use tools I've used for years that I never discovered. It analyses code-bases faster than I can grep, and interrogates command-line tools to understand their operation in real-time.
Automation Everything
Setting up repositories, managing commits, creating CI workflows, generating documentation - AI can automate these tasks by actually running the commands and learning from the output.
Real Examples That Matter
E-ink Display Debugging
I asked AI to SSH into our target board, capture webcam images of the code running on our controller chip, analyse what was displayed, identify mistakes, and iterate until the e-ink display showed correctly. This normally takes days - we did it in hours.
Automated Power Optimisation
We're now automating access to test equipment in our remote hardware lab. AI monitors power consumption as we make BSP changes, takes automated readings, and feeds results back into the optimization process for better battery life.
The Critical Gotchas
- Content Overload: AI creates far too much content and makes things unnecessarily complex
- Rabbit Holes: It gets confused and can chase irrelevant solutions
- Happy Compliance: Ask it to do something nonsensical and it will absolutely do it
- Tracking Management: You need to keep it focused on what you actually need to achieve
The Essential Insight
This only works because I stay in control. AI is incredibly powerful, but you need:
- Your experience to know what questions to ask
- Feedback loops like CI testing and target board validation
- Constant course correction to prevent hallucinations
Community Discussion
How do we share these techniques? What feedback loops work best? How do we prevent AI from creating technical debt while harnessing its genuine productivity gains? I've got real examples to share, but I want to hear what's working for others.
Alex Lennon has been working with embedded Linux and OpenEmbedded for at least 15 years (possibly longer, but his memory is shot). He's the founder of Dynamic Devices Ltd and delivered a keynote at Yocto Project Summit 2022.
Over the past few months, Alex has been experimenting with AI-assisted development workflows for Yocto projects, learning what works, what doesn't, and where the real productivity gains lie. His recent work includes automated hardware testing, remote target debugging, and power optimization workflows - all enhanced by AI collaboration.
Alex believes the embedded community needs to figure out AI integration together, sharing both the successes and the failures to develop practical best practices.