However you found your way here, welcome — and thank you for being curious enough to read. If you didn't arrive through Agent Fang, the live app this newsletter is built around, you can open it here; it is the fastest way to see what I mean below.
I have built three of these AI agents on AWS, and until now I have kept them mostly to myself. There is a Chinese saying — bi-men-zao-ju, building a car behind a closed door. It never works; you open the door and the car doesn't fit the road. So it is time to open the door. I have been a full-time developer for sixteen years, building software that has to survive inside real company infrastructure. The last stretch of that has been agents.
Everyone can ask ChatGPT how to build a fireplace in the backyard. The harder question is the one at work. You already Google less and ask Claude or ChatGPT more when in doubt — that is the floor now, the minimum. The real test is what comes after. Would you connect your work email to a model and let it sort and reply for you? Would you install a plugin into your live database and let it run the analysis? That is where comfort runs out, and that gap is where this newsletter lives.
I have watched that gap up close. An internal AI department stood up an offshore team to build a full-stack solution and push it to users. The users ignored it — they would rather reach the AI through the tools they already knew. The solution worked. Nobody used it. That was a company big enough to write off a quarter on software nobody opened. A five-person shop has no such slack: the cost of the unused tool lands on the same people it was supposed to free up. Same gap, less room to absorb it. The industry is excited about how powerful the tools are, while the rest of the world is still working out how to actually use them.
What I am putting in front of you is Agent Fang. On the surface it is a flow engine with a few AI elements — the tip of a floating iceberg. Underneath is a pipeline: Claude Code and Codex inside VS Code, through GitHub, out to AWS, run by a team of AI assistants, each playing a different role. I'll write about that team in a later edition. One honest caveat: Agent Fang is a single-shot flow player, not strictly an agent — it does not loop back and start over at the end. I built it that way on purpose. Watch the chips along the top; each one is a phase, and you can see the state change as execution moves through them. To see is to believe.
After Codex and Claude Code shipped, the number of apps released each month on Apple's App Store doubled. Attention is the scarce thing now, and flashy demos compete for a slot on your phone and your workstation. How many of them get installed and never opened again? At work that mistake costs far more. A proof-of-concept demo used to be the thing that convinced people, and thanks to AI the cost to stand one up has dropped hard — cheap enough that I can build one on request, a message away. But that is the trap: a demo proves something can be built, not that it will get used. My day job is answering technical questions as a senior developer, and every time I do, I ask the same thing — how is this person going to use it once I walk away? The honest answers never show up in the demo. They show up in the long stretch after it, when the novelty wears off and the thing has to hold up at work. That is why this newsletter is called Post Demo: the view from the back office, after the applause.
Agent Fang is also an argument about how this software gets built. The default path is to hire a consultant or an offshore team and wait. Agent Fang is what that process looks like when you replace the consultant with an agent-augmented build — run by one developer who keeps asking how you will actually use the result. I built it for myself first. The open question, the one I cannot answer from my own desk, is whether the same approach holds inside your business.
That is the ask — and not the generic "let me know what you think." Pick one of these and reply in two lines:
What is one task at work you would hand to an AI agent, and what has stopped you so far?
Where does your comfort run out — email, your database, customer-facing replies, something else?
Have you had an AI tool pushed on you that you ended up routing around? What did you use instead?
If I built a demo for your exact situation, what would it have to do to convince you?
I read every reply, and over time they shape where this newsletter goes — more than my own guesses about what you want ever could.
"You can have a superior product, but if it doesn't fit into somebody's workflow, if it doesn't fit into their day, it's tough to get adoption."
— Cameron Davies, head of AI at Yum! Brands, The Economist, January 2026