What Have We Been Up To?

It has been months since the last post, and a lot has changed. Let me try to catch you up.

First, my free time went very sporadic for a while, and then I left my position at Amazon. I’m now working on Arbor full-time, trying to polish it enough to offer as a product.

Second, my family welcomed a new addition in July. My adopted daughter’s baby sister has joined us and is blessing us with many late-night feedings, diaper changes, crying spells, and smiles that make it all worthwhile.

Third, I turned my focus away from Arbor’s agent features to build out its software factory. My intuition was that I could accelerate development of the agent features once I automated more of the software engineering work.

This has turned out to be both true and false at the same time. The factory can now build most features with minimal human intervention, but it builds more slowly than I did by “hand” because every piece of AI-generated code has to clear new quality gates (a 10 perspective LLM council) before it lands. That’s the part I’m currently fine-tuning, and the balance shifts with every new model release.

I don’t consider the slowdown a failure. I consider it the actual cost of AI engineering. But it’s a big enough topic that it deserves its own post.

Anyone who has been building in AI over the last few years can attest that the only constant is change. Really, that describes everything in life, not just tech. But the recent velocity has been something else, with “new” terms and best practices hitting the headlines and making everyone feel like they’re getting left behind if they don’t adopt immediately.

Frankly, my take is that most of it is the hype cycle doing hype cycle things. A partial list, roughly in the order they showed up:

• Prompt Engineering
• Retrieval-Augmented Generation (RAG)
• Tool Use / Function Calling
• Agentic AI
• Model Context Protocol (MCP)
• Context Engineering
• Agent Memory
• Agent Skills
• Agent Evals
• OpenClaw / Harness Engineering
• Tokenmaxxing
• Loop Engineering
• World Models

I might not have them in exactly chronological order, but you get the gist.

Most of that list is noise and rebranding to keep the hype levels high. The ones that actually deserve to be kept are maybe three:

1. Prompt Engineering — though it matters less now that the AI can write good prompts itself.
2. Context Engineering — a lot of the list collapses into this rather than standing on its own.
3. Agent Evals — the only way to know whether any of the rest is working.

Put those three together and you’re basically building a harness. Loops and tokenmaxxing are memes that should never have entered our vocabulary.

• We have a prompt library that has become an agent skill library. Agent Skills are a great place to organize your reusable prompts.
• Agent context is heavily engineered, with memory and evals feeding self-improvement.
• Arbor has held a model-agnostic stance since the beginning, on the expectation that models keep improving and the design should absorb those improvements as they arrive.
• Security and accountability have been largely ignored by the mainstream, so Arbor has invested heavily there. I’m a security engineer by trade, and I wanted to lean into that from the start and build something secure by default. I want Arbor to be safe and simple enough for my mom to use, and also something an enterprise could actually run in-house. (I’ve written about where we’ve fallen short of that too.)

I’ve spent so much time building tools that work for me that I’ve neglected the new user experience. The goal is to get Arbor into a releasable state where a new user can paste a single command into their terminal and get a local instance running. I want to hit that by the time ElixirConf rolls around in September.

I’m also building Arbor to be voice-first. This is a risky move, but it’s what I need for myself. I’ve played with web chat interfaces, a TUI, a Signal integration, and more. But the pace of my life right now means I need something hands-free. I often have a 17-month-old in one arm and an infant in the other. Typing becomes a real trick in that configuration.

To be clear, voice-first is not voice-only. Earbuds are the primary interface, but the phone is the visual one, and anything that actually needs review goes there as text, such as diffs, plans, long lists, anything sensitive. The spoken side just tells you it’s waiting. Every voice interaction is transcribed and stored. And approvals for anything destructive can’t happen by voice at all; those escalate to the screen where you can read them before you accept.

That doesn’t mean the other interfaces go away — it means they’re second-class citizens unless someone joins the project to mature or rebuild them. And yes, I realize that’s a major point of friction for anyone considering Arbor as their daily driver. I’m betting it pays off in the long run. Think J.A.R.V.I.S., with a phone screen for graphical output instead of a heads-up display, and you’re halfway there.

That’s all for now. I’m still very excited to be building in the current age of AI. We’re all trying to keep pace with new tech — many of us independently making and learning from the same mistakes — but I see the future as bright and full of opportunity.

If you’re going to be at ElixirConf in September, come find me. And if the voice-first bet strikes you as brilliant or as a mistake, I’d like to hear either one.

Until next time!