The AI Coding Revolution Has a Huge Problem?

A few weeks ago, I stumbled across a debate that has been making the rounds in software engineering circles.The spark came from Boris Cherny, an engineer at Anthropic and the creator of Claude Code, arguably the most influential AI Agentic Coding harness in the world today.During a podcast appearance, Boris made a statement that immediately grabbed my attention:Coding is largely a solved problem.He went on to explain that he hadn’t written a line of code by hand since November, and that essentially all of his code is now authored by Claude Code.Needless to say, this generated strong reactions.Some people interpreted it as evidence that software engineering is about to be fully automated. Others saw it as confirmation that AI coding tools are delivering unprecedented productivity gains.As someone with twenty-five years of experience in software development, I found myself somewhere in the middle.Because while I absolutely believe AI is transforming software engineering, my own experiences suggest that programming is nowhere close to being a solved problem.In fact, the more I use these tools, the more complicated the situation appears.The Rise of Agentic Software DevelopmentOver the past few years, we’ve witnessed a rapid evolution in how software gets built.First, developers used AI to generate snippets of code.Then they began using AI assistants to complete larger programming tasks.Today, we’re entering what many people call Agentic Software Development.Instead of asking an AI for a few lines of code, developers increasingly delegate entire workflows.The AI can analyze requirements.Generate designs.Write code.Create tests.Review its own output.Deploy software.Monitor production systems.In theory, the human becomes less of a programmer and more of an orchestrator.The promise is obvious.If AI agents can perform most of the implementation work, then software engineers can become dramatically more productive.Ten times more productive, according to some advocates.Perhaps even more.At least, that’s the dream.My First Encounter With Enterprise Agentic AIIn 2025, I returned to work after a lengthy medical leave.My wife had experienced a serious health crisis, and I had spent months focused almost entirely on family.When I came back, one of the first things I noticed was that my employer had become obsessed with Agentic AI.Leadership had heard about tools like Claude Code and Cursor.They had heard stories about developers becoming ten times more productive.Naturally, they concluded that our company needed its own internal version.Let’s call it Kevin.Kevin was our homegrown agentic harness.Compared to Claude Code, Kevin felt slower, heavier, and burdened with enterprise compliance guardrails.It ran on older-generation models.It often struggled with context.And yet, despite all its flaws, Kevin was still capable of orchestrating significant portions of software development.Using Kevin gave me a front-row seat to what Agentic AI actually looks like inside a large enterprise.What I observed left me both impressed and concerned.The Business Knowledge ProblemOne of the biggest weaknesses I encountered had nothing to do with coding itself.It had to do with understanding.AI models are remarkably good at generating software.What they are not particularly good at is understanding the business context behind that software.Organizations often operate on thousands of unwritten assumptions.Knowledge exists in hallway conversations.Slack threads.Meeting notes.Institutional memory.The heads of senior employees.Much of this information never appears in formal documentation.Humans navigate these gaps naturally.AI agents do not.If a requirement is not explicitly documented, the AI will often substitute something that appears reasonable based on its training data.The generated code may compile successfully.The unit tests may pass.The architecture may look elegant.And yet the implementation may completely miss the actual business objective.This problem becomes particularly severe in large brownfield systems where decades of accumulated business logic exist beneath the surface.The Context Window WallAnother challenge is context.Every AI model has a limited working memory.For small projects, this isn’t a major issue.For enterprise software systems containing millions of lines of code, it becomes a constant battle.Once an AI agent exceeds its effective context window, strange things begin to happen.The model forgets previous decisions.It ha