2026-04-02
By Vadym · Generated with Boba, curated by me
The developer role is evolving through 5 distinct eras — from writing every line to governing agent fleets. Most teams are stuck at Era 2 (AI-assisted). The hardest jump is to Era 3 (AI-native), where you write specs instead of code. The transition is organizational, not technological.
April 2026 · Explore the full interactive research at evolution.suprun.ca
Two years ago, I wrote every line of code by hand. Today, I write specifications and review what agents produce. The shift happened faster than I expected, and if you're a working developer, it probably happened faster than you expected too.
I've spent the last several months researching how the software engineering role is transforming — reading DORA reports, McKinsey studies, GitHub's Octoverse data, practitioner accounts from Steve Yegge and Addy Osmani, and enterprise adoption frameworks from AWS, Microsoft, and PwC. What emerged is a pattern that keeps repeating across every source: this isn't a single change. It's a staged evolution, and most of us are somewhere in the middle of it.
I'm calling it the 5 Eras of the AI Developer.
Era 1: The Traditional Engineer (pre-2022). You write every line. The IDE, Stack Overflow, and documentation are your tools. Only 14% of professional developers still work this way, according to GitHub's 2025 Octoverse report — but the foundational skills this era built (debugging intuition, systems thinking, deep language knowledge) remain essential at every stage that follows.
Era 2: The AI-Assisted Engineer (2022-2023). You code with GitHub Copilot or similar tools, accepting or rejecting suggestions. You're still the primary author — you understand every line — but AI accelerates the routine work. The ratio is roughly 60% manual, 40% AI. This is where most organizations are today: 84% of developers use AI coding tools regularly, and Copilot alone has 20 million users deployed across 90% of Fortune 100 companies.
Era 3: The AI-Native Engineer (2024-2025). AI does most of the implementation. You write specifications; agents write code. Your primary activities become defining requirements, reviewing AI output, making architecture decisions, and maintaining quality gates. The ratio flips: roughly 20% manual, 80% AI. The industry term that's winning is "agentic engineering" — spec-first, accountable, disciplined.
Era 4: The Agent Orchestrator (2025-2026, emerging now). You coordinate multiple AI agents working in parallel. One handles the API layer, another the UI, a third writes tests. You define system boundaries, manage agent workflows, and verify outputs at scale. The ratio: about 5% manual, 95% AI. Cursor's CEO disclosed that 35% of their internally merged pull requests are created by autonomous agents.
Era 5: The System Governor (2027+, predicted). Your primary product is system design and constraints, not implementation. You define "what" and "why" — never "how." Fleet-scale agent systems with AI supervisors coordinating groups of coding agents are the tool landscape. This era is largely theoretical, but it's already visible at the frontier.
The identity shift tells the story most clearly: Author, Editor, Director, Orchestrator, Governor. Each era changes not just what you do, but who you are in the development process.
Explore the interactive version: The full research — all 5 eras, the data, transition playbooks, and failure modes — is available at evolution.suprun.ca.
Most engineers I talk to are somewhere on the Era 2-to-3 boundary. They use Copilot or a similar tool daily. They accept suggestions, generate boilerplate, and move faster on routine tasks. But they haven't made the deeper shift to spec-driven, agent-delegated work.
This is the hardest transition in the framework, and the research backs that up.
The Era 1-to-2 move is straightforward: install a tool, use it for two weeks on non-critical work, build intuition. GitHub's own rollout template gives it 90 days. But the 2-to-3 transition requires something fundamentally different. It's not a tool change — it's a workflow redesign.
Here's what the data shows about what happens when teams make this jump:
The bottleneck shifts. DORA's 2025 report found that code review time increased 91% and PR sizes grew 154% in teams using AI heavily. Implementation gets faster, but the review queue gets longer. You're not saving time — you're moving the constraint.
Specification writing becomes the new premium skill. It was never a formal discipline in traditional development. Now it's the primary bottleneck. Teams that deploy agentic tools without training developers to write effective specs, prompts, and acceptance criteria get generic output and conclude "AI doesn't work for our codebase."
Quality doesn't improve automatically. Bug rates climbed 9% even as velocity increased (DORA 2025). AI-co-authored code showed 1.7x more major issues and 2.74x higher security vulnerabilities in one large-scale analysis. The METR study found experienced developers took 19% longer with AI tools — but believed they were 20% faster. The perception gap is real.
Organizations matter more than tools. As DORA put it: "AI magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones." Only 32% of organizations have formal AI governance policies. McKinsey found that only 30-50% of pilot gains translate to business impact without workflow redesign.
I went through this transition myself. I moved from Copilot to Windsurf to Kiro to full agentic workflows with CLI-based coding agents over about eight months. The tools changed, but the real shift was in how I think about work. I spend more time writing specifications than writing code. I spend more time reviewing than implementing. The skills that matter now — decomposing problems into clear, testable units; writing acceptance criteria that an agent can execute against; knowing when to trust output and when to intervene — are fundamentally different from the skills that made me effective five years ago.
I'm experimenting with Era 4 now — running multiple agents in parallel on isolated workstreams. It works, but it's not magic. The coordination overhead is real. Managing three agents on clearly separated tasks (API, frontend, tests) is manageable. Managing five on interconnected work creates more overhead than it saves.
The pattern that works: decompose into truly independent workstreams, define clear interfaces between them, and establish quality gates where no agent output merges without passing automated tests plus human review. It's less like coding and more like managing a small team of very fast, very literal junior developers who never get tired but also never ask clarifying questions.
The companies furthest along this path are instructive. Cursor reports 35% of merged PRs from autonomous agents. Teams are shrinking from 6-10 people to 3-5 while producing 2-5x output. But trust remains fragile — one survey found trust in agentic systems dropped 89% between May and July 2025.
Three things are clear from the research:
The transition is not optional. 80% of the engineering workforce will need to upskill through 2027 (Gartner). 75% of jobs require redesign by 2030 (McKinsey). The market is already pricing in Era 3/4 capabilities — entry-level tech hiring decreased 25% year-over-year in 2024, and internship postings are down 30% since 2023.
Speed without discipline is dangerous. The organizations that skip governance, review processes, and cultural work see quality regression. Klarna cut 40% of staff, replaced roles with AI, saw customer satisfaction drop, and had to rehire. The lesson: "AI as cost reduction" is a trap. "AI as capability amplifier" is sustainable.
The skills that matter are shifting upward. Kent Beck said it well: "The implementation stuff has gotten easier, and that is exactly why other skills like breaking down projects, setting goals, and project management become much more highly leveraged." The premium is moving from typing speed to judgment.
I built an interactive research experience that maps the full framework — all 5 eras, the transition playbooks, the data, the failure modes. If any of this resonates, start there.
And ask yourself honestly: which era are you in?