When I started building Arcjet in early 2023, the landscape was the same it had been for a decade. Code was written by hand – probably with Copilot autocomplete – over months of iteration. VS Code. Pull requests. CI pipelines. AWS or one of the newer serverless clouds. ChatGPT had launched, but it was just a chatbot. The challenge was finding good engineers who could bring experience to something novel.
Three years later, everything is different. We’re past the inflection point (Opus 4.5 & Codex 5.2). Building software is no longer the hard part – it never really was.
The three-stage development cycle – planning, coding, testing – is compressing into a single continuous loop. Code is written by AI agents in hours or days, which means the review bottleneck has shifted from syntax to architecture and judgment. Multiple workstreams run in parallel, orchestrated by one developer. Everything engineers historically deprioritized – comprehensive tests, accurate documentation, architectural decision records – now directly determines whether the AI produces quality output or slop. The editor is being replaced by the agentic terminal.
At Arcjet, I’ve shipped multiple major features in recent weeks, alone, that would have taken months of engineering a year ago. Everyone on the team who has fully embraced this shift is seeing the same thing.
Building is no longer the bottleneck. The complaint about low-quality AI code doesn’t hold up unless you think most human-written software was ever well written. Quality used to be scarce – most teams couldn’t ship reliable, well-tested, well-documented software fast enough to matter. An experienced developer using AI tools well can now ship more functionality, with better test coverage, more reliably documented, in a fraction of the time.
The pace is intense. New model capabilities, new interfaces, new protocols – each becomes essential, then gets superseded. The engineers who will survive are the ones who treat this as the job now: directing agents, double-checking the riskier areas, maintaining the context that makes AI useful. The job description is changing faster than most people are updating their mental model of it and those who aren’t on board will find it increasingly hard to be hired.
Running the console.dev newsletter for developers, I now receive tens of new devtool submissions every day, up from maybe a couple each week just at the beginning of 2026. Every developer with an AI subscription and a weekend can now ship something credible. The result is a Cambrian explosion of software – like the internet, cloud, mobile, and crypto eras all compressed from decades into months, with every founder in the world paying attention simultaneously.
Now it’s about knowing what to build, how to stand out, and how to get it in front of customers. That was always at least half of the battle of a startup. Now it’s most of it.
A small team with genuine domain insight and an existing audience has never been better positioned – because the ability to complete no longer hinges solely on how many features you can build. The engineering gap between a two-person startup and a hundred-person org has collapsed. What remains – judgment about what to build, relationships, trust, reputation – can’t be replicated with a subscription. If you have those things, the opportunity is infinite.