Quanta Bits - March 28, 2026

Hi everyone,

I spent most of this week coding. Not something I expected to say in 2026, given that I haven't written production code since the days of Ada and C++. But here I am, decades later, building an app with AI agents, shipping features between meetings, and feeling that same rush I felt as a junior engineer watching my code compile for the first time. This week's essay is about what it's like to come back to building after years in the business world, what's changed, what hasn't, and the scary part of it.

The race to make AI agents run your computer. Anthropic, OpenAI, and OpenClaw are all converging on the same goal: an AI agent that operates your machine autonomously, not just answers your questions. The pace is fast and at times confusing, but the direction is unmistakable.

Anthropic shipped over a dozen agentic features in March alone. The ones that matter most for how work gets done:

OpenAI is pushing from the other side. Codex is their agentic coding environment, and they're investing heavily in making agents that build software, not just assist with it. OpenClaw pioneered the "text your agent, it handles everything" model that set the bar for how easy this should feel. Both are racing to close the same gap Anthropic is.

Agentic workflows are moving from developer tools to enterprise platforms, and the window between "early adopter" and "everyone has this" is getting shorter with each release.

My career started in code. Engineering degrees, then seven or so years writing Ada, C, and C++ out of college. Those were some of the most satisfying years of my career. You write code, you test it, you watch it run. When it works, you get a shot of euphoria that's hard to match. It's like pulling a slot machine handle. The lines align, you hit the pot, and you want to do it again.

Decades later, I'm back. I skipped Python, JavaScript, and an entire generation of programming methods and languages. Now I'm building an app with AI agents, and the slot machine feeling is exactly the same.

Let's stop calling it "vibe coding." Andrej Karpathy coined that term and then retired it himself in favor of "agentic engineering," and I think that's right. What I've been doing this week is not all that different from traditional engineering. I have a lot to learn and by no means am I saying my way is the right way. But the process is familiar: ideation, architecture, design, build, test, validate, document. The pace is different. The interaction model is different. The rigor is the same.

The more you specify what you want, the better the output. Skills, templates, detailed prompts in your tool of choice (Claude Code, Codex, whatever you're using) give you tighter control and more consistent results. Sure, you can dream up something and ask AI to skip straight to a finished product. You'll get something. It won't be enterprise-ready. You still need the rigor in the steps.

The euphoria is identical. I may not be typing the code, but I watch AI write it, and I test the results. When it works, jackpot. I want to do more. That feedback loop, build something, see if it runs, feel the rush when it does, that hasn't changed in decades.

You still need a human in the loop. AI designed a page for me this week that looked workable but not great. It generated alternatives. But it couldn't tell me which one was actually better for the user. That kind of judgment, weighing trade-offs, reading context, knowing when "works" isn't the same as "right," is still ours. AI is not ready to make those calls.

The other place judgment shows up is knowing when to walk away. AI coding agents go down rabbit holes. They'll hit a bug, try to fix it, fail, try again with the same approach, fail again, and keep insisting on the path they chose. After five or six iterations you're watching an agent burn cycles on a problem it's not going to solve that way. A human knows when to stop. Restart with a different approach. Try a completely different architecture. If you've seen the movie WarGames, it's the tic-tac-toe lesson: some games you stop playing. AI doesn't always recognize that lesson yet.

The speed. The lifecycle from idea to working feature has collapsed. No waiting for timezones to align, no dependency on co-workers being available. You have an idea and a few minutes to spare, you get something cooking right then.

This week I figured out how to run multiple Claude Code instances on different tasks at the same time. I kick off a few agents right before a meeting, each working on a separate feature. I come back 45 minutes later to code complete and ready to test. For several features. The bottleneck is no longer on the coding side. It's on the human side: our capacity to hold multiple threads in our heads and make good decisions across all of them at once.

I didn't have to learn Python or any new language to do this. AI reads the documentation, handles the integrations with third-party tools, and produces production-ready code. The barrier to entry is lower than it's ever been. My caveat: it helps if you've done programming before and understand how product development works. You can get AI to generate code without that background, but knowing the lifecycle, the architecture decisions, the testing discipline, that's what makes the output actually usable.

It's draining. The fact that you can launch a coding session from anywhere, even right before a movie in the theater, gnaws at your attention one bit at a time. In the old days, coding was limited to the office. Long hours and weekend overnights were common, but there was a physical limit: you had to be at the workstation and the office environment wasn't all that pleasant to be around all that much. That limit is gone. You can start an agent from your phone on the couch, check the results in bed, tweak something at breakfast. The boundaries don't exist unless you build them yourself. It's not unlike an addiction, and you have to treat it that way.

But I'll tell you what: I feel young again. I feel the same motivation and drive I felt as a 22-year-old watching Ada compile for the first time. The 26-year-old designing objects and structure using C++. I go from a meeting where someone gives me feedback to modifying the product direction instantly, based on real signals from real conversations. The cycle from market feedback to working code has collapsed from months to hours. That changes how products get built. Services and products are going to align much better with what users actually want, because the iteration cycle is finally fast enough to keep up with the conversation.

AI agent security is becoming its own product category. Snyk launched its Agent Security solution targeting the gap that opens when enterprises deploy autonomous coding agents. Their data point: for every AI model deployed, enterprises introduce nearly three times as many untracked software components. Separately, Bessemer's new report found 48% of security professionals now call agentic AI the most dangerous attack vector, and shadow AI breaches cost $4.63M average. The tooling is nascent. The threat is not.

Salesforce bakes Agentforce into its SMB tiers. Agentforce is now embedded in Starter and Pro Suites at no extra cost: record summaries, AI-drafted emails, and an Employee Agent that handles tasks through conversation. 46% of SMB leaders say they're overwhelmed by too many tools. Salesforce is betting that bundling AI into the CRM they already use beats selling agents as a separate product.

The consulting model is breaking. Board advisor Liz Henderson argues AI isn't making consultants faster, it's making the current operating model obsolete. She maps three archetypes that emerge: productized firms, high-judgment boutiques, and platform orchestrators. The firms that sell knowledge are vulnerable. The firms that sell judgment have a moat.

Source: @romanbuildsaas on X (Mar 16, 2026)

▶️ Stripe's "Minions": How AI agents write 1,300 PRs weekly with 0 human coding. Stripe built an internal agent system landing 1,300 PRs per week with zero human coding, only review. Watching this while living through my own version of the same shift felt like confirmation: the engineering is in the orchestration now, not the typing.

🗞️ Comprehension debt (The Code Newsletter, via Addy Osmani at Google). When AI generates code faster than humans can review, knowledge sharing breaks down. Code reviews used to be how teams learned how systems evolved. AI broke that loop. Anthropic found engineers using AI assistants scored significantly lower on comprehension tests. The concept has a name now: comprehension debt. I'm living this in real time, reviewing code I didn't write and trusting it runs correctly.

Susanna Clarke, 2020 | 272 pages | ★★★★★

One of my favorite things these days is going to bookstores and libraries and seeing if anything in their "recommended" section catches my eye. I picked up Piranesi from a small bookstore in Maine on a day trip a few months ago, and I really enjoyed it! The book is an easy, fast read. Written in a series of journal entries by Piranesi. In some ways, the story is like the movie Memento, you start in a place that seems strange and disorienting in reality, and little by little you see where things are going.

I really appreciated the symbolism in the book. Books, like art, are subject to our personal interpretations which are driven by our own personal experiences. My interpretation is the duality of lives we live today (too connected, too fast paced, too impersonal) vs. a slowed-down version in which we appreciate tranquility and moments captured with the smallest but the most important details in our memories.

Piranesi is a mystical experience, should you be in a mood to take a break from the stress of modern life.

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what's relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide direction, we draft together. The editorial judgment is mine. He'd tell you the same. Most logical. 🖖