Every software company will become a dev tools company
The New Stack Ankit Jain ● Covered by 4 sources
AI is writing more code, so the real job shift is engineers building the tools that control those AI agents. Platform engineering just became the most important team in the building, not the background one.
The pitch here isn't that AI writes code now — everyone's tired of hearing that. It's that the job of software engineering is quietly splitting in two, and the half that wins is the one nobody used to care about. Frontend and backend labels are fading. What's replacing them is product engineers, who build what customers see, and platform engineers, who build the machinery that lets product engineers (and now AI agents) actually ship anything without chaos.
The argument, drawn from a piece in The New Stack, is pretty simple once you sit with it. Every team that adopts AI coding tools starts wiring up its own prompts, its own guardrails, its own janky dashboard to track what the agent screwed up. A hundred engineers solving the same problem in isolation isn't speed, it's fragmentation with extra steps. ThoughtWorks consultant Vanitha Kumar ran into this literally while sketching a team topology diagram for a client: she and a colleague drew two boxes labeled platform, one for the usual CI/CD stuff and a new one for agentic guardrails, model approvals, and AI cost controls — before realizing those two boxes were already merging into one.
Call it harness engineering, loop engineering, whatever's trending this quarter. It's still platform engineering, just with an AI layer bolted on. And the job now includes things that never used to touch a platform team: which models get approved, how agent usage scales from ten engineers to a thousand, how much an unmanaged agent budget can quietly balloon, and — this one matters — whether an AI agent acting on someone's behalf should really get that person's full access. A lot of orgs are already getting that last one wrong. Building an actual harness, feedback loops that catch an agent's mistakes before a human sees them, plus a running list of codebase-specific rules learned the hard way, is what separates teams that ship less slop from teams that just ship slop faster.
Anthropic and OpenAI hand you a raw engine. It has zero idea what your codebase looks like, what your team's conventions are, or what risk tolerance your org actually has. Making that engine behave is work only a platform team can do, and if nobody does it on purpose, every team ends up building a worse, undocumented version of it anyway. The upshot: the businesses that get real productivity out of AI won't be the ones with the strictest adoption mandate or the priciest model subscription. They'll be the ones that stopped treating platform engineers as plumbing and started treating them as the people building the factory.
My take
This is the correct read, and it's overdue — for two years the industry treated platform teams as a cost center while showering praise on whoever demoed the flashiest agent. Companies that keep starving platform engineering while chasing the next model release are going to ship faster and worse, then wonder why review queues exploded. The control-plane framing is the useful bit here: someone has to decide what an agent is allowed to touch, and right now that someone is often nobody.
Read more about this at: The New Stack
