Looking back at 2025, one theme dominates: AI agentic programming went from novelty to necessity. This year transformed how I think about software development, career advice, and even my personal hardware projects. Here’s what I wrote about, what I learned, and what I expect in 2026.
The Big Picture: AI Agents Arrived
The year started with tentative exploration and ended with full embrace. In April, I declared that Agentic Programming: The New Robot Revolution! was upon us. MCP (Model Context Protocol) had just launched, A2A (Agent-to-Agent) specification emerged, and I noted that improvement cycles were happening every two months - compared to Moore’s Law’s two years. I called agents “ephemeral robots” - no motors or sensors, but definitely wired to the real world.
By December, I was writing that Coding is Solved (And That’s Fine) - the mechanical act of translating requirements into syntax is increasingly handled by AI. What still matters: understanding code at a gears-level, foundational CS knowledge, and software “taste.”
Claude Code Became My Primary Tool
My journey with Claude Code evolved dramatically through the year:
July: In From Clicking Yes to Letting Claude Run Wild (Safely), I discovered VS Code devcontainers as the perfect sandbox. The constant “mother may I” permission dance was tedious, so I found a way to give Claude full permissions inside a controlled container where the worst that can happen is blowing away the container.
November: Using Claude Code to Build gocat showed the power of documentation-first AI development. I ported the RFCat Python library to Go for the YardStick One RF transceiver. The key insight: mount reference code read-only so Claude can learn from it but can’t modify it. The result? 100% packet success rate at 64 packets per second.
December: Putting Claude in Container Jail: My localdev Setup introduced my localdev project - a Podman-based development environment specifically designed for running Claude Code in “dangerous mode.” Git became the real safety net, and I developed the mantra: “commit early, commit often, commit constantly.”
The Go Advocacy Arc
A strong thread through the year was my growing conviction that Go is underappreciated for production systems:
In Why Go for Robotics?, I questioned the Python-for-prototyping-C++-for-production pipeline. Python’s hidden costs (GIL, dependency hell, cross-compile complexity) and C++’s complexity tax both have alternatives. Go offers single-binary deployment, excellent cross-compilation, and goroutines that make concurrent programming almost easy.
The year ended with Programming Languages in 2026 - a five-factor analysis of language choice: popularity, jobs, LLM proficiency, concurrency, and deployment. Go wins two factors (concurrency and deployment), making it the strongest all-around choice for production services.
Career Advice in the AI Age
Advice to New Developers (or, How to Get Hired) was blunt: the job market is brutal, companies are in a holding pattern waiting to see how AI transforms their businesses, and if AI makes you even 1.5x more productive, companies will need 50% fewer developers.
My guidance: master AI-assisted development, become an agent architecture expert (MCP!), focus on building things rather than submitting applications, and break the stereotype of engineers who can’t communicate. The 10x engineer of 2025 manages thousands of agents.
Hardware and Radio Projects
The year started with a personal milestone: Finally Got my Ham Radio License - and Whales! I became K06HAX, a key step toward building a “telephone system to the whales” - AI on a buoy with a transducer and radio link to shore.
I also documented practical maker knowledge: Pulseview on Linux for logic analyzer setup and Agent Building: Ollama Hosted Models where I tested local LLM performance (spoiler: the M4 Pro MacBook crushed it at 43 TPS; the Ryzen AI 9 limped at 17 TPS because ROCm doesn’t support gfx1150 GPUs).
Addressing the Skeptics
AI Agent Naysayers tackled the critics head-on. Many haven’t seriously tried AI pair programming. Yes, agents make mistakes. Yes, critical code needs extra care. But “trust, but verify” is what we do anyway. The economics are undeniable, and today’s models are the worst you’ll ever use - they only get better from here.
I reviewed Vibe Coding by Gene Kim and Steve Yegge - two skeptics who became true believers. Their conversion story matters more than any marketing pitch. The core message: experienced engineers can treat AI as a capable junior developer who needs clear direction and verification.
A Personal Note
And yes, I also wrote about Foreign Cinema - my favorite restaurant in the world, in San Francisco’s Mission District. Some things are more important than code.
What I Expect in 2026
Based on everything I learned and wrote this year:
-
AI coding agents will become table stakes. Companies not using them will fall behind. The productivity gap will become undeniable.
-
Go will continue rising for production services. Its combination of deployment simplicity, excellent concurrency, and growing LLM support makes it ideal for the AI-augmented workflow.
-
The junior developer role will transform. They won’t learn the way we did. Mentorship becomes critical. Universities need to radically change how they teach software engineering.
-
Local models will improve but lag. Running Ollama locally is fun but cloud APIs will dominate for serious work due to performance gaps.
-
Container isolation for AI tools will become standard. Nobody wants to run
--dangerously-skip-permissionson their actual system. -
Agent swarms are coming. A2A enables specialized agents collaborating on complex tasks. One for architecture, one for security reviews, one for tests, one for documentation.
-
The skill premium shifts from “writing code” to “system thinking, troubleshooting, and clear written communication.”
The Bottom Line
2025 was the year I went from AI-curious to AI-dependent. Not blindly trusting - always verifying - but genuinely dependent. The tools are here. They work. They’ll only get better.
The question isn’t whether to adopt AI-assisted development. It’s how quickly you can effectively integrate it into your workflow.
The agent swarm is coming. I’ll keep writing about it as I go.
This blog post was written with AI assistance. Of course it was.