The same demo is everywhere right now: an agent running for hours, untouched, shipping a whole feature on its own. Nobody is demoing the discipline. The hype is outpacing it: the loop is autonomous, the verification isn't. Look at the conditions the demo runs under. The token budget is effectively unlimited, of course it is, the vendor...

The risk with agentic coding isn't bad code. It's a generation of engineers shipping code they don't understand. The fix isn't a course. It's building the learning into the delivery itself. Because what you lose to agents was never the ability to produce code, the agent produces it. It's judgment: the mental model of why the code is sh...

Everyone's braced for agents to ship code fast and sloppy. I just reviewed one that shipped fast and better (see attached example) and the reason has nothing to do with how smart the model is. Look at what it does before it asks for your time. It tells you what it changed and, just as usefully, what it deliberately left untouched. It p...

Amazon has a leadership principle for this: invent and simplify. 37signals has a name for the second half: the wind tunnel test. AI has made the first half easy. Nobody has solved the second. You can ship a working prototype in days now. Test the idea, get real signal, move on. That part is basically solved. But "we shipped it" and "we...

"All checks have passed. This branch has no conflicts with the base branch." That green box is now the trigger some teams use to let AI agents merge their own pull requests. No human in the loop. Green means go. The reaction tends to split the room, half excitement, half horror. I understand the excitement. If the agent can write the c...

Can pair programming embed peer review directly into the development workflow? And could that be an answer to the PR review bottleneck AI-assisted delivery has created? Some teams think so. I've heard of pairing used exactly that way to combat review bottlenecks. Not because two people type faster than one. They don't. Because the revi...

For over a decade, Ruby on Rails was the only stack I fully trusted myself in as a solo or small-team engineer. This year, AI made that loyalty optional, and Python is the proof: I shipped my first production code in a language I'd never used professionally, and trusted it. Claude Code reviewed that first Python code and said the craft...

The adapter pattern is one of the first design patterns every engineer and agent should understand — and every integration you’ve shipped, from Stripe to a maps API, is why. The bigger question now is whether your AI coding agent knows what it is, and when to reach for it. Almost every external-provider dependency starts the same way: ...

Use TDD. Refactor carefully. Follow SOLID. Use design patterns. Write clean code. I've seen agent instruction files full of phrases like this. The problem isn't that any of it is wrong. It's that none of it is self-executing. An AI coding agent mostly remixes what it already finds in the codebase. If the surrounding code has clear boun...

AI code review tools like CodeRabbit and Greptile are useful. But if the plan is to let the agent build it and then let the reviewer clean it up afterwards, you may already be too late. By the time a review tool sees the PR, the agent has already made the decisions that matter. It has chosen the shape of the implementation, made assump...

The people getting the most out of AI agents are not the best prompters. They are the best operators. Keep the human at the start of the loop and at the end of it. Let the agent own the middle. That is the whole job. Prompting is where everyone starts. Ask a sharper question, add more context, tweak the wording, try again. It works, an...

The fix for AI code review isn't more reviewers. It's making the code worth reviewing before it ever reaches one. Addy Osmani (Director of Cloud AI at Google) just wrote a sharp piece on this, and the numbers in it are brutal. Review duration up 441%. Zero-review merges up 31%. Defect rates climbing from 9% to 54% (Faros AI). Roughly 4...

Spec-first AI development is quietly rebuilding the one thing engineers spent a decade escaping: waterfall. The agent doing the typing won't save you from it — if anything, it makes the trap worse. That's why I'm sceptical of approaches like Spec Kit becoming the default way teams work with coding agents. Not because specs are bad. Spe...

Your next customer might not be human. On parts of the internet automated traffic already outnumbers people, yet most businesses still build their entire online buying experience for someone who is. Cloudflare has reported that automated traffic now overtakes human requests in parts of the internet. HUMAN Security found AI-driven traff...

AI can write more pull requests than your team can confidently review. Governed PRs are how you keep both — the quality bar and agentic speed. More PRs only help if review confidence keeps up. Otherwise every reviewer starts from scratch, asking the same questions before they can trust a single line. Did the agent follow our standards?...

AI code review asks whether the code looks okay. Governed delivery asks whether the change is acceptable. That distinction matters more every week, as more teams move to AI-assisted and agentic development. AI review tools are genuinely useful. They spot issues, suggest improvements, catch the obvious mistakes, and take some repetitive...

A year ago the message to engineers using AI was: use the agents fully, lean in, let them run. The ones burning the most tokens were read as the ones embracing the future — and rewarded for it. The cautious ones, the engineers asking what all this was costing, got quietly filed under not forward-thinking enough. Now the message is flip...

AI's immediate value is not automation. It is amplification. Most businesses have not figured that out yet, and it is costing them. They are asking "how do we automate this completely?" That question will matter eventually. In some areas it already does. But right now, for most real businesses — software delivery, operations, finance, ...

Too often it means more process, more forms, more meetings, more gates, and more delay. Red tape dressed up as control. That is not what I’m trying to build with Software Dark Factory. The point of SDF is not ceremony. It is to help teams keep useful delivery moving while making AI-assisted work easier to review, trace, and challenge. ...

Old delivery had a slow feedback loop. Friction built up over months. Slow test suites, review bottlenecks, unclear ownership, repeated governance steps. Someone eventually noticed and maybe it got fixed. With AI-assisted delivery, that loop is too slow. The system moves faster, which means waste compounds faster. And the cost isn't ju...

AI-assisted software delivery changes how teams measure the cost of engineering work. That sounds like a finance sentence, but I do not think this is mainly a finance problem. It is a delivery problem. For a long time, the cost of a software change was mostly hidden inside human effort. How long did it take to understand the request? H...

That is the thought that kept coming back as I tried to work out what AI coding agents should change about software delivery. The pressure right now is real: move faster, ship more, use the agents properly, let the tools take more of the work. I understand the pressure. I also understand the appeal. But a lot of what gets described as ...

More engineering teams are talking openly about agents auto-merging PRs. The reaction is predictable — half excitement, half horror. The excitement is understandable. If your AI agent can write the code, pass the tests, and merge the PR without anyone waiting in a queue, that is a genuinely faster delivery loop. The speed case is real....

Your team is shipping faster. Your review queue is longer. That’s not a coincidence. This is the pattern I keep seeing from engineering leaders, founders, and developers who have adopted AI-assisted development seriously: the output speed is real, but the delivery model hasn’t caught up. PR review backs up. Rework increases. Decisions ...

You have written this line. So have I. def can_access? stripe_subscription_id.present? endIt looked fine at the time. Probably week three of the project. Billing was working, Stripe was sending webhooks, and access was gating correctly. Ship it. That line is not a billing integration. It is a governance failure you have deferred. Here ...

Specs are useful. They are not governance. Spec Kit exists because vibe coding made an old software problem suddenly more dangerous. Vague intent has always produced bad software. The difference now is speed and plausibility. A loose prompt no longer produces only a rough prototype. It can produce a feature-shaped object with files, te...

Once I had the engineering side of the Dark Factory model working, governed pull requests, agentic workflows with clear handoffs, playbooks applied on every PR, a different problem became obvious. Engineering was touching marketing surfaces. And marketing had no equivalent governance layer. That is not a complaint about marketing. It i...

Governance is not something I learned from a book. It is something I could not afford to get wrong. Twenty years of working in startups does something to the way you think about shipping software. Not because startups are glamorous or fast or any of the things people say. But because in a small team, there is nobody to hand the respons...

Building the engineering side of the Dark Factory operating model was easier than I expected, and the reason was clear in retrospect: I had 20 years of SDLC experience to draw on. I already knew what a good pull request workflow looked like. I knew what a reasonable CI/CD pipeline should do, how code review should work, what a deployme...

AI learns the same way a junior engineer does. Through reinforcement. Through feedback. Through someone watching closely enough to catch the drift early and correct it before it becomes a habit. Most people do not think about it that way. They treat an agent like a search engine with ambition — give it a task, expect a result, get frus...