AI Coding

Technical debt lives in your code. Cognitive debt lives in your head. Intent debt lives in the artifacts you never wrote - the goals, constraints and rationale that explain why the system is the way it is. It’s the one kind of debt your agents can’t pay down for you, and the one agentic engineering makes most expensive.

As we dug deeper, however, a different problem emerged. No one on the team could explain why certain design decisions had been made, or how different parts of the system were supposed to work together. The code might have been messy, but the deeper issue was that the team’s shared understanding, the theory of the system, had quietly fragmented. They had also failed to write down or communicate the rationale behind decisions. They had accumulated cognitive and intent debt faster than technical debt, and it had paralyzed them.

SymJack is a new attack technique targeting AI coding agents: a booby-trapped repository to trick your AI coding assistant into overwriting its own configuration through a disguised file copy, then run attacker code on the next restart. This is one technique that works against the whole category, don’t treat it as six separate bugs.

The era of fixed-fee AI spending just ended. If you’re a CTO or engineering leader and you haven’t noticed yet, you will very soon — probably around September 2026 when some budget alerts start firing.

There’s a specific code smell that shows up in AI-generated code, and once you see it you can’t un-see it: primitive obsession all the way down to the domain core. string for emails. string for IDs. Map<string, any> whenever the situation gets hairy. Working code, passing tests, ships fine. And yet a developer who’d actually thought about the domain would not have written a single one of those types that way.

Cognitive offloading is delegating to the AI and still owning the answer. Cognitive surrender is when the AI’s output quietly becomes your output and there is nothing you feel is left to check. For software engineers the line between the two moves under your feet most days, and most of us are crossing it without noticing.

An analogy came to my mind this winter: vibecoding is fast food. You can cook at home (write the code yourself), you can go to a good restaurant (hire a programmer), or you can buy yourself a hamburger (use AI tools).

Over the past 16 months, DX has been running a longitudinal study on AI’s impact on engineering velocity across a sample of more than 400 engineering organizations. We found that as AI tool usage increased by an average of 65%, median PR throughput increased by just under 8%. Most organizations are landing in the 5–15% range—a meaningful gain, but far below the 3x or 10x expectations many leaders are being held to.

Comments

Comments

Comments

More agents running doesn’t mean more of you available. The narrative around agentic engineering right now is still mostly about throughput and parallelism, but almost nobody’s talking about what it actually costs the human in the loop.

A well-built outer harness serves two goals: it increases the probability that the agent gets it right in the first place, and it provides a feedback loop that self-corrects as many issues as possible before they even reach human eyes. Ultimately it should reduce the review toil and increase the system quality, all with the added benefit of fewer wasted tokens along the way.

Two releases this week, however, suggest we’re starting to enter a second phase of open source agentic coding projects. The first brought us *clones*, this next phase brings us *reimaginings*.

Claude Code went from 0.7% to 4.5% of all public GitHub commits in six months

Comments

Here’s what I keep seeing: the specification-driven development (SDD) conversation has exploded. The internet is overflowing with people saying you should write a spec before prompting. Describe the behavior you want. Define the constraints. Give the agent guardrails. Good advice. I often follow it myself.

You’re paying for the AI to write code, so surely the expensive part is what it produces, right? I would have guessed something like 60/40 or maybe 70/30 skewed towards output. The reality is nothing like that. Real-world data from OpenRouter’s programming category shows 93.4% input tokens, 2.5% reasoning tokens, and just 4.0% output tokens. It’s almost entirely input.

It turns out Claude Opus 4.6 and GPT-5.4 are both very competent at SwiftUI—and a full SwiftUI app can fit in a single text file, which means I can use them to spin something up without even opening Xcode.

The biggest risk in AI adoption isn’t moving too slowly. It’s mandating too hard, watching it backfire, and oscillating between full commitment and full retreat. The organizations that found their footing got better at paying attention, not better at writing policy.

AI conversations are ephemeral by design — decisions made early fade as sessions lengthen, and nothing survives the session boundary. Developers hold on to long conversations not because long sessions are productive, but because the context lives nowhere else. I propose externalizing decision context into a living document — external memory that persists what the context window cannot, turning transient alignment into durable shared understanding.

My efforts rewriting machina and postal over 13 days — not to mention the ~8 other libraries I’ve written in the last year — have served as proof that stranded expertise is much cheaper to rescue than ever.

It’s like a Trojan Horse of craft: guidelines you might never agree to explicitly are guiding LLM outputs, which means you are agreeing to them implicitly.

A recent piece, Grief and the AI Split, made me ask myself what exactly it is that I like about building software.

Comprehension debt is the hidden cost to human intelligence and memory resulting from excessive reliance on AI and automation. For engineers, it applies most to agentic engineering.

After building Curling IO Version 3 in Gleam alongside AI coding agents, I’m convinced Gleam is the best language for this workflow. Agents don’t write better Gleam - there’s less training data. But Gleam’s compiler lets agents self-correct without waiting for a human.

TL;DR: AI-assisted coding is revealing a split among developers that was always there but invisible when we all worked the same way. I’ve felt the grief too—but mine resolved differently than I expected, and I think that says something about what kind of developer I’ve been all along.

AI doesn’t fix your bottlenecks. It amplifies them. DORA’s report on AI-assisted software development confirmed that AI amplifies existing flow and exposes brittle processes and frictions. If your flow is good, AI makes it better. But if you have a lot of painful bottlenecks and frictions, AI makes the pain worse.