How Software Developer Skills Are Changing in the Agentic Coding Era

TL;DR: Between late 2025 and early 2026, tools like Claude Code reached a capability threshold that radically changed developer workflows. The key roles are now Software Engineers and Architects, while traditional junior developer, language specialist, and manual tester positions are declining. Strategic competencies — systems thinking, problem decomposition, critical judgment — become the most valuable because they are the hardest to automate.

An Epochal Turning Point

The tipping point has arrived, evident, tangible, unequivocal.

It had been in the air for a long time, the signs were clear, and yet it was surprising.

Personally, I had been waiting for this moment since 1990, the year of my degree. My thesis was on Artificial Intelligence: “A Prototype Expert System for Statistical Conceptual Design”.

A system that received specifications in natural language and produced a conceptual database model.

The work was part of a larger project at the Department of Computer Science and Systems of Sapienza University of Rome, whose ultimate goal was to create an AI tool that would allow software to be developed using Italian instead of any programming language.

Those were times of excitement about AI. Japan had funded a major national project to reach the Fifth Generation of Computers. Truly intelligent computers capable of passing the Turing Test, something like HAL 9000, similar to Star Trek computers that activated at the phrase: “Hey, Computer!”.

The Fifth Generation was expected by 1993, 1995 at the latest.

The Fifth Generation of Computers never arrived. In its place came yet another long AI winter.

In more recent springlike times, some awaited AGI, Artificial General Intelligence — it’s more or less the same thing, but perhaps that won’t arrive soon either.

But perhaps it doesn’t matter, because in the meantime, just when the enthusiasm around LLMs was beginning to cool, something else arrived. Something of more limited scope, but equally useful and powerful.

The scope is precisely that of my thesis: Software Engineering.

Between late 2025 and early 2026, several models — Anthropic’s Opus 4.6, OpenAI’s GPT-Codex-5.3, and Google’s Gemini 3 — reached a capability threshold that radically changed developer workflows. Boris Cherny, creator of Claude Code, stated that in his last month as an engineer he didn’t open an IDE: all committed code (about 200 PRs) was written autonomously by AI. According to a senior Google engineer, Claude Code can accomplish in one hour what would take a year.

This is not a media stunt. It’s all true.

Between July 2025 and the publication date of this article, at Rome’s Digital Innovation Hub, companies like Chirale and Fastal collectively made about a dozen major software releases, without any developer writing a single line of code.

The New Operational Reality

Rakuten reports testing Claude Code on codebases of 12.5 million lines, completing complex tasks in seven hours of autonomous work with 99.9% accuracy.

TELUS reports creating over 13,000 customized AI solutions, shipping code 30% faster and saving over 500,000 hours.

Another important element concerns code quality.

Code developed with the support of generative AI, up until spring 2025 — the product of pair programming activities assisted by traditional ChatBot applications or tools like GitHub Copilot — did not have the characteristics to be considered production ready.

From June 2025, everything changed.

Tools like Claude Code are far more complex than simple apps like ChatGPT or Claude.ai.

As I wrote on the Fastal blog, these are products based on a new architectural paradigm, capable of harnessing LLMs and focusing their behavior with a strong emphasis on the specific project.

The output of these agents are high-quality codebases, coherent and compliant with their respective project standards, complete with all those implementation details that make the difference in terms of quality, such as accurate exception handling, strong typing, schema validation, and automated test suites.

True science fiction for many domestic software houses.

Within the Fastal Group, the qualitative difference between newly produced codebases and the ocean of legacy code was so evident that we launched refactoring and consolidation processes so extensive they saturated the generous usage limits of our high-tier subscriptions to Anthropic & Co.

Layers upon layers of technical debt, accumulated on our oldest systems, were swept away in a few months, with incredibly low effort.

All of this, however, is not automatic.

These results are only possible through a profound and radical reorganization of software development work and a careful reskilling process for senior developers.

A New Mix of Professional Roles

The traditional mix of professional roles, based on a few senior engineers or architects and many developers, must change.

The traditional developer role is transforming: in 2026, the task is no longer writing code but directing AI agents that write it.

Key competencies include problem decomposition, formulating precise prompts, reviewing outputs, and the ability to reason about systems, failure modes, and architectural trade-offs.

As observed by Andrej Karpathy, former AI Director at Tesla and ex-OpenAI, there exists “a new programmable layer of abstraction” that includes: agents, sub-agents, prompts, contexts, memory, tools, plugins, and IDE integrations.

Software Engineers and Architects

The key roles become Software Engineers and Software Architects, professionals capable of having a systemic vision of applications and projects, able to evaluate and discuss solutions, technology stacks, and fundamental implementation choices.

In the new organization of the agentic coding era, they are the human factor that governs and empowers the harnesses to which the more operational work is delegated.

Basic competencies in technologies and design patterns are not sufficient. Special soft skills and context engineering capabilities are needed.

AI Product Manager

Product managers are acquiring an AI-enhanced role: competitive analyses that previously required weeks now complete in hours, and vibe coding allows rapid prototyping of ideas without depending on engineering. The 2026 PM must know how to translate product vision into technical specifications that AI agents can effectively implement.

Quality Assurance

Manual testing as we have known it is becoming obsolete. AI agents can generate functional and exploratory tests automatically, predict edge cases, and keep test suites in step with software evolution.

QA professionals transform from executors to strategists: they supervise test agents, define coverage strategies, and focus on creative testing, UX, and ethical considerations.

Declining Professional Roles

UX Designer

The traditional app development cycle, based on wireframe prototyping, mockups, Figma rendering, and hand-off to developers, is no longer compatible with the coding speed available through new tools.

The UX Designer becomes an unacceptable bottleneck, bringing the effective speed of project advancement back to traditional levels.

Tools like Claude Code can develop fully functional UI prototypes in minutes, autonomously launch interface execution, capture screenshots, and evaluate — leveraging the computer vision capabilities of the SOTA LLMs that power them — compliance with expressed requirements as well as aesthetics and usability best practices.

In this regard, the interview with Jenny Wen, former Figma and current Director of Design at Anthropic, for Lenny’s Podcast is particularly significant, with a very clear title: “The design process is dead. Here’s what’s replacing it”.

Traditional Junior Developer

The most obvious victim.

The maturation of coding agents in the second half of 2025 destroyed the lowest rung of the professional ladder.

Writing boilerplate, generating unit tests, and refactoring legacy code — activities that trained juniors — are now at marginal cost zero.

An Indeed report reveals that software development roles decreased by 3.5% year-over-year, with a particularly steep decline in entry-level positions.

78% of tech roles in job postings require familiarity with AI.

This creates a critical paradox: if companies eliminate junior positions, there will be no one to fill senior positions in the future. The risk is a “Competence Cliff” — a competency chasm in the talent pipeline.

Language / Specific Stack Specialist

The distinction between “Java developer”, “PHP developer”, or “frontend/backend developer” is dissolving. With AI writing competent code in nearly all popular languages and frameworks, a backend engineer can generate decent frontend, cross-platform, or even native mobile code through prompts.

Why search for and hire separate frontend and backend developers when a harness like Claude Code expresses itself perfectly in any modern programming language and applies the best design patterns, in a perfectly idiomatic manner, in any modern framework without problems and delays due to the learning curve?

Software architects in 2026 can finally choose the technology stack best suited to the project, without constraints tied to the company’s existing skill portfolio.

Competencies in frameworks and languages shift from idiomatic skill and the quantity of syntactic knowledge to the ability to critically evaluate the correct application domains of solutions.

Traditional Manual Tester

Positions based exclusively on manual testing are increasingly rare. Agentic AI can simulate human interactions, continuously test in production-like environments, and provide actionable insights faster than any human.

The shift is from a “manual-first” QA model to an “automation + AI-first” model.

Ticket Implementer

The developer who receives a well-defined JIRA ticket and implements it without asking questions is a dying breed.

Today, companies pioneering AI automation, such as Cursor, already have automations that automatically forward tickets to AI agents, which handle debugging, fixing, and code deployment.

The Strategic and Cross-Cutting Skills That Matter

“Soft” skills paradoxically become the hardest to automate and therefore the most valuable:

Systems thinking: the ability to reason about architectures, failure modes, and trade-offs is the true differentiator between those who direct AI effectively and those who don’t.
Problem decomposition: defining problems clearly, establishing precise constraints and requirements is the competency that determines AI output quality.
Domain expertise: deep knowledge of the business domain enables asking AI the right things. Without understanding of maintainability, security, and scalability, AI won’t include them spontaneously.
Critical judgment and ownership: AI produces code rapidly, but someone must take responsibility for correctness, reliability, and consequences.
Cross-functional collaboration: the boundaries between PM, designer, engineer, and QA blur; the ability to work across functions becomes essential.

Impact on Large Enterprise Organizations

The revolution brought by new tools doesn’t only concern software houses. The introduction of AI, when well orchestrated, can make a difference even in more structured companies operating in any industry.

From Functional Hierarchies to Platforms and Pods

Organizations rethought in an AI-first logic must abandon traditional functional hierarchies in favor of platform-based structures, with modular and reconfigurable cross-functional teams, organized around data flows and algorithmic processes.

Instead of departmental silos — marketing, operations, finance — work must be structured in “pods” or “squads” that combine technical specialists, domain experts, and operational personnel with shared end-to-end responsibility.

Smaller Teams, Greater Output

AI’s multiplier effect enables smaller teams to produce output previously reserved for much larger teams. Processes that required two days can now be completed in hours.

All those ad hoc automations that couldn’t become outsourced software development projects can today be solved with throwaway app generation.

This implies a profound restructuring of IT department staffing: fewer generalist developers, more senior figures with orchestration capabilities and strategic vision.

The Talent Pipeline Paradox

This scenario, however, reveals critical issues.

The most insidious question is the training of new generations. If manual coding has historically been the training ground for junior developers, eliminating these positions means eliminating the formative pathway that produces tomorrow’s seniors.

As in the old adage “writing is thinking”, so too “coding is thinking”: when everything is delegated to AI, there is a risk of losing the deep understanding of the problems that code attempts to solve.

A balance must be found: at Fastal, we are proposing AI as a tool for accelerated learning (not as a substitute for learning) and redesigning junior career paths around supervision, review, and systems thinking competencies from day one.

Recommendations for Software Industry Companies

Within our Trade Union at CNA Roma, we are defining a set of recommendations for companies in software development:

Redefine hiring profiles around architectural capabilities, systems thinking, and domain expertise, rather than mastery of specific languages
Invest in AI-generated code governance: automated tests, robust CI/CD pipelines, and security standards become critical infrastructure
Maintain a redesigned junior talent pipeline: new hires must learn to direct AI, not be replaced by it
Adopt platform-based organizational structures with cross-functional pods and algorithmic governance roles

We also have recommendations for young talents entering this sector now:

Develop architectural vision: it’s the most powerful multiplier in the AI era. Knowledge of design patterns, optimization, security, and technical debt management is no longer limited by typing speed
Master agent orchestration: understand how to structure prompts, contexts, permissions, and multi-agent workflows
Deepen business domain knowledge: AI amplifies whatever domain expertise you possess — robust testing, modular architecture, and CI/CD practices make AI safer and more powerful
Never stop understanding code: even if you no longer write it manually, the ability to read, evaluate, and debug remains fundamental to ensuring quality and security

For a deeper look at agentic coding tools, read my article on Claude Code and the era of agentic coding.