Videos

Jacob Beningo explores the shift from using AI as a coding assistant to agentic firmware development, where AI agents autonomously build, debug, test, and verify embedded systems. He shares a real experiment using an NXP MCX N947 board and an AI-written I2C driver. He also covers hardware safety guardrails, simulation, and how embedded engineers can evolve into systems engineers.

Chapters:
00:00 Introduction: From Coding Assistants to Agentic Firmware
02:35 Why AI Skeptics Are Wrong: Workflows and Work Queues
04:20 Giving AI Iteration Through Hardware and Simulation
05:58 Experiment: NXP MCX N947 and an AI-Written I2C Sensor Driver
08:55 The Backwards Click Board and a Production-Grade Driver
09:49 The Agentic Firmware Workbench: Skills, Agents, RAG, and MCP
10:27 Giving Agents Observability: Debuggers, Scopes, and Power Supplies
11:40 Why Agentic AI Helps Time-Starved Teams and the Junior Engineer Concern
13:28 Risks of Hardware Access and the Role of Simulation
15:16 Boundary Conditions and Safety Guardrails in MCP Servers
18:14 Human Oversight, Verification, and Moving Beyond Low-Level Drivers
20:34 How to Start: Automate One Repetitive Task at a Time
22:00 Closing Thoughts, Resources, and Newsletter

Key Takeaways:
• Most teams use AI only as a coding assistant. The bigger gains come from agentic workflows where AI builds, debugs, tests, and verifies firmware.
• AI doesn't get firmware right on the first try, and neither do human engineers. It needs the ability to iterate, as with test-driven development.
• Connecting an AI agent to real hardware or a simulator lets it observe results, find root causes, and fix problems instead of guessing.
• In the I2C experiment, the agent correctly diagnosed a hardware problem (a backwards click board) that the engineer had wrongly blamed on the software.
• Once the board was fixed, the agent produced a tested, production-grade temperature and humidity sensor driver within about an hour.
• A robust agentic workbench combines skills, agents, RAG, and MCP servers, with an orchestration agent following your own development processes.
• Give agents access to debuggers, oscilloscopes, logic analyzers, and adjustable power supplies. This brings real risk, including fire hazards, hardware damage, and liability.
• Build safety into MCP servers. For example, require users to set voltage and current limits up front, and prompt for approval whenever the agent tries to exceed them.
• Simulation offers a safe, continuous way to validate AI-generated code, especially when hardware is dangerous to leave unattended.
• Start small. Pick a repetitive daily task with a clear outcome, automate it, then expand the workflow gradually while keeping human verification and static analysis in place.