Donn Felker, AI platform lead at Polygon, formerly of Tinder and an Android specialist for ten years, publishes an essay on September 14, 2026 about the disappearance of software engineering as he practiced it. He dates his realization to July 2025, based on three facts. He built a complete SaaS without writing a line of code. At Tinder, a two-thousand-word prompt executed in ten minutes, without error, a threading library migration that used to take six to eight hours per instance. Finally, he saw non-technical people ship real products: a CFO built his financial operations tool with the Claude app, a gym manager his coaching platform. His conclusion: it was never about coding but about executing — code was just a toll.
The essay then explains engineers' anxiety through two social psychology concepts. Occupational categorization reduces a person to their job; professional identity centrality makes that job the foundation of their worth. When day-to-day work shifts to an agent, identity wavers: either fall behind or train one's own replacement.
Felker proposes the image of a leverage pendulum between product and engineering. The agent shifts the initial leverage toward product, which gets a prototype in one afternoon. But complexity and scale pull the leverage back toward engineering: a CFO stuck after two months in the same Claude Cowork session without knowing where his code lives, a photographer whose Base44 app stored his media on Google Drive at an unsustainable transfer cost. The engineer fixes it, puts in place architecture and guardrails, then hands back control. Each oscillation dampens, until the next model version.
The role that emerges at the equilibrium point is the agentic builder, a hybrid of engineer and product lead. Pure specialists at both extremes are presented as under threat. The proposed path has two steps: codify one's knowledge into skills and procedures executable by agents, which makes the individual scalable and shifts them from operator to orchestrator; then adopt systems thinking — business for the engineer, engineering for the non-technical person — with a list of six requirements to put to the agent: architecture, guardrails, backups, versioning, governance, observability.
The author sets a six-month horizon for making the transition and concludes with the figure of the builder able to read a balance sheet and an execution trace.
Key takeaways
Three triggers dated to July 2025. a complete SaaS built without writing code; a 2,000-word prompt — "which we'd now call a Skill" — that executes in 10 minutes a threading library migration that cost an engineer 6 to 8 hours per instance, repeated hundreds of times across Tinder's codebase; non-technical people shipping products. The story of this shift echoes that of [[cherny-sequoia-coding-is-solved-loops-printing-press-2026-05]].
Central formula."Coding was just a gatekeeper." What separates dreamers from doers is the capacity to execute; code was only ever the toll for it.
The identity framework. occupational categorization compresses a person into a role ("Android Developer"); professional identity centrality makes that role the foundation of perceived worth. Hence the double bind the essay formulates: not using AI means falling behind; using it means training one's own replacement.
The leverage pendulum. before AI, engineering set the pace and held the leverage; the agent shifts the initial leverage toward product (an MVP in one afternoon). Complexity, scale, and systems economics pull it back toward engineering, which fixes things, puts in place architecture and guardrails, then hands it back to product. Each return is shorter — damped oscillation — but a new model version restarts the motion.
Two cases where the pendulum swings back. a CFO who built his financial tool in a single Claude Cowork session over two months, until the environment reached saturation, without knowing where his code lives (found in a private GitHub repository); a photographer who wired Google Drive up as media storage in a Base44 app, a choice the agent followed without objection and that made transfer costs unsustainable. The agent "will follow your lead".
The coach analogy. just moving is enough at first, then you plateau or get hurt; the engineer corrects form and sets the program, then steps back; subsequent visits are shorter.
The role that emerges. neither engineer nor product, but both — "the engineer who thinks like a product person and the product person who thinks like an engineer". On the engineering side, the proposed title is Agentic Product Engineer. Pure specialists at both extremes are presented as exposed; deep specialists persist but with skills outside their domain. This mapping of roles can be read alongside [[ng-ai-engineering-skills-map-2026-08-14]].
Two-step path, common to both profiles. (1) codify one's knowledge into skills, templates, and procedures executable by humans or agents — "you're making yourself scalable", a shift from operator to orchestrator; (2) move from specialist thinking to systems thinking, which the author calls cognitive complexity.
For the engineer. letting go of an identity tied to a language or platform (the author hasn't touched Android or Kotlin in six months), treating each skill as a Lego brick, weighing build vs buy with maintenance in mind, thinking in terms of ROI, acquisition, retention. A position close to [[osmani-google-new-sdlc-vibe-coding-agentic-engineering-2026-05]]'s on the shift to agentic engineering.
For the non-technical person. ask the agent at every blocker, and demand six things they wouldn't know to ask for on their own — maintainable architecture, guardrails, backups, version control, governance, observability. Recommends installing a Destructive Command Guard.
⚠️ The six-month timeline is repeated three times without justification; it functions as a rhetorical device, not a substantiated forecast.
Recommended first step. list what you do manually and dread doing, pick one thing, ask Claude or ChatGPT "help me turn this into a skill", and let yourself be interviewed. Five minutes.
Key figures
threading library migration at Tinder reduced from 6-8 hours to 10 minutes with a 2,000-word prompt, July 2025