When Everyone Can Code: The Deep Split Transforming Software Engineering

Recent blogs and articles reveal that AI can already write code, but the more consequential shift is in the software engineering profession itself. The authors observe a clear division of the field into at least two emerging disciplines.

Harness Engineering focuses on constructing constraints, tools, feedback loops, and documentation that make AI agents operate reliably. Charlie Guo’s "Harness Engineering" playbook illustrates this trend, with OpenAI reorganizing teams around agents, Stripe’s internal AI system "Minions" merging over 1,000 pull requests weekly, and independent developers running five to ten agents in parallel without reading the generated code line‑by‑line.

Judgment Manufacturing cultivates humans who can guide, verify, and maintain the output of agents, especially early‑career engineers. While agents lower the barrier for anyone to perform Harness Engineering, deep engineering skills remain indispensable for safety, reliability, performance, compliance, and handling high‑stakes scenarios.

The authors stress that a system built through Harness Engineering can be functional, but it cannot circumvent physical laws, resist attacks, or manage the entropy of production environments without human judgment.

Combining Guo’s article with Greg Brockman’s public remarks, the authors distill a reproducible engineering pattern:

Agent‑first by default – start with 5‑10 bullet points of requirements, let the agent draft a solution and pull request, then review and approve.

Architecture as guardrails – enforce strict boundaries and allowed dependency paths, automating structural checks.

Tools as foundation + feedback – expose internal tools via CLI/MCP so CI, lint, and test failures are precise enough for agents to self‑repair.

Memory that compounds – treat AGENTS.md as a repository of “scar tissue”, recording failures and mitigations for cross‑team reuse.

Plan‑first discipline – require agents to produce a plan, undergo review, obtain approval, then execute.

No slop policy – maintain merge standards; every pull request must have a human owner who fully understands and signs off.

Agent ops layer – manage agents like production systems, tracking runs, centralizing tool permissions, and turning recurring failures into harness improvements.

These practices constitute both workflow hygiene and executable infrastructure.

The article then examines the explosion of bespoke software . Andrej Karpathy recounts building a highly personalized cardio dashboard with an agent in about an hour, highlighting that the traditional App Store model—choosing from discrete apps—is becoming obsolete. The future lies in AI‑native services, sensors, and actuators that assemble on‑demand into “ephemeral software”.

Andrew Ng adds an economic perspective: even a ten‑fold increase in developer productivity does not reduce the need for developers because the ceiling for custom‑software demand is effectively limitless. He notes the emergence of “X Engineer” roles (e.g., Marketing Engineer, Recruiting Engineer) embedded within business units.

Collectively, these observations point to a macro trend: software is shifting from packaged products to a continuous, on‑demand custom tool flow, possibly marking a transition beyond “Software 3.0”.

Thomas Wolf (Hugging Face) expands on the decreasing cost of rewriting code. When rewriting and understanding external codebases become cheap, deep dependency trees lose their strategic advantage, and “replacement pain” diminishes, weakening the Lindy effect. Agents can extract needed parts or rewrite entire modules, reducing attack surface and package size.

However, validation remains a bottleneck. The authors quote Wolf: “unknown unknowns remain unknown,” emphasizing that while agents accelerate output, they do not proportionally accelerate confidence. Formal verification, therefore, shifts from optional to mandatory in an AI‑dominated world.

Data on the junior talent pipeline underscores a strategic risk: from 2022 to 2026, entry‑level developer job postings fell 67 % and hires dropped 73 %; unemployment rates for recent CS graduates rose to 7.5 %, surpassing many non‑technical fields. The “Junior Death Spiral” describes how demand for junior roles has plummeted, with many positions now requiring three or more years of experience.

Russinovich and Hanselman argue that AI coding assistants amplify senior engineers—who possess the judgment to harness and verify agent output—while potentially hindering early‑career developers. Their proposed remedy is large‑scale preceptorship: pairing early‑in‑career engineers with seasoned mentors at a 3:1 to 5:1 ratio for over a year, and embedding an “EiC mode” in AI copilots that uses Socratic coaching instead of direct answers.

Looking ahead, the authors predict three developments in the next twelve months:

Harness Engineering will solidify as an official job title, similar to the evolution of DevOps.

Bespoke software will become mainstream, increasing demand for engineers who can operate, integrate, and secure these custom solutions.

The erosion of the junior pipeline will become a strategic risk, as teams that stop hiring and training junior talent gain short‑term output but accrue long‑term fragility.

Ultimately, agents are turning software into an abundant business; the scarce resource will be human judgment—determining what is good, proving it works, and keeping it working.