This week is the one-year anniversary of the last time I typed code, though I’ve been producing software full time this entire year. The gradual increased adoption of AI in my coding was so subtle that I hadn’t realized until late Dec 2025 that I had last coded by hand in October that year.
My year of pure agentic coding hasn’t been anywhere near as productive as Michael Novati’s 4M lines of code YTD, but it has been equally dramatic from a life-change perspective. If it wasn’t clear to people a year ago, it should now be blindingly obvious no one will ever be paid to handcraft code again. Doing so will be as rare, and as pedagogically motivated, as handcrafting assembly.
The last year has seen not only dramatic improvements to coding agents and harnesses, but significant qualitative changes, perhaps permanent, to the profession of software development.
I made a roundly ridiculed claim seven months ago that software companies would move to lights-out codebases where human intervention would not only be unnecessary, but both impracticable and ultimately dangerous.
Though many, especially wizened cynical senior developers, still criticize my claim around the inevitable takeover of light-out codebases where humans don’t review code, none of the premises of my argument have changed:
As AI gets better at coding, humans subtract when intervening
AI will be better than people at reviewing code, if it isn’t already
It’s impracticable for humans to review the volume of AI generated code
We’re not at a point yet where humans touching code merely introduces errors… but anyone coding full-time this past year has undoubtedly experienced many cases where AI either independently tackles edge cases their human pilots didn’t anticipate or highlights problematic lines in PRs which their human authors missed.
Just as humans are error-prone, AI-generated code will continue to have errors. But these curves have already intersected. It’s clear AI coding agents supplemented by AI review agents will surpass human coding teams in quality if they haven’t already.
Not just impractical, but impracticable — that is, “utterly undoable.” Naysayers of my original argument never addressed this (to me) undeniable deduction:
I used to spend about an hour most mornings at Meta reviewing others’ code
Engineers now typically report 5-20x productivity gains using AI to generate code
… thus it’s mathematically impossible for a Meta employee to review every single line of code
This argument isn’t even about the quality of the code produced. I’m making a quantitative, hours-in-a-day argument. Anybody who claims their team still hand-reviews every line of code is either:
working on a team that’s shockingly inefficient at piloting AI, or
glossing over PRs, hopping amongst lines already highlighted by review bots
Anyone extrapolating will easily see where this is going. Calvin Grunewald and Vivek Sharma just launched ArchDev this month, one of many products which assist with the impossibility of hand-reviewing code.
Lights-out codebases aren’t anywhere near as ridiculous as they may have sounded a mere seven months ago. I bet this will increasingly be acknowledged as inevitable over the next year.
The profession of software engineering has experienced more change in the past year than it likely ever has in history. People who’ve coded several decades no doubt feel a mix of excitement and whiplash — as well as, if you’re like me, a bit of sadness that something we enjoyed so much has so quickly become relegated to a hobby you’ll never again be paid to do (i.e. handcrafting code).
Over three years ago, when I was part of a tiny minority using Copilot within VS Code, AI mostly tab-completed enums and did hilariously bad things, like the one time Copilot inserted some random guy’s email to my comment block (George something, IIRC).
But I’ve now spent a year full-time typing English prose into Claude Code and Cursor. My workday feels qualitatively different in a few ways:
I’ve embraced speech recognition (Aqua and Wispr Flow) as a faster way to prompt. More than speed, I find that the effortlessness of speaking allows me to issue fuller prompts with more context, yielding better results.
I feel the exact opposite of “flow,” instead micro-sharding my attention amongst myriad conversation threads. When agents are busy, I flip to email/Slack/Messages to fill the idle time. This is both highly productive and deeply unsatisfying. By the end of my work days, I feel scattered and anxious.
I work more than ever. As with most productivity tools promised to us as time-savers, the ability to use a tool in more scenarios simply means we work more. I’m now disappointed when flights don’t have WiFi because I need to keep my agents unblocked. I often don’t head to sleep until I’ve issued a few time-consuming instructions that can be executed overnight.
Agents know just enough to be annoying. PR review bots have become the nit-picky pedant who calls out massively unlikely edge cases requiring preternatural architectural gymnastics to address. If you leave PR bots to iterate on their own with coding agents that fix all their suggestions, you will soon watch the codebase devolve into a nightmare of gratuitously fragile complexity.
I miss working with people. The personal empowerment has been amazing. Until AI, I had not realized how much of my professional work weeks were spent simply trying to convince coworkers to go along with an idea. Now I just implement whatever I want, and quickly. But the best thing I did the past year was to collaborate regularly with a former coworker, Ashok Kuppusamy, to brainstorm and build together several hours each week. I’m calling it now: the future won’t be solopreneurs, but duos and trios of people building together.
The ability to build tools cheaply has led to a large Batman Utility Belt of one-off tools which save me tons of time. For instance, I’ve got a skill that compares versions of the same podcast episode which differ in their dynamically-inserted ads; the skill downloads, transcribes, and chapterizes those different versions, then analyzes whether ad-skipping has worked correctly for a specific customer of Superphonic. This sort of analysis used to take me 20 minutes each. Now I just fork an agent each time this comes up. The ability to cheaply build custom tools will likely mean our work experiences will start to greatly differ from each other as every person crafts their surroundings with tools to their own liking.
It’s been a crazy year. I can’t wait to see what the next one brings.