AI Reshapes R&D: From Skill-Driven to Intent-Driven Paradigm

Part 1: Coding Evolution – Four Stages of AI Autonomy

The article outlines a four-stage progression of AI autonomy in coding:

Stage 1 – Manual Coding: Traditional fully manual development.

Stage 2 – Copilot Coding (Dec 2024 – Jan 2025): AI assists with local code generation, improving efficiency at the node level.

Stage 3 – Vibe Coding (Jan 2025 – Dec 2025): Natural language guides AI to generate, modify, and debug code; AI can independently complete small requirements.

Stage 4 – Spec / No Coding (from Jan 2026): AI autonomously fulfills requirements end-to-end based on specifications; humans focus on business understanding and architecture design.

Across these stages, three dimensions evolve simultaneously: AI shifts from assistant tool to collaborative partner to autonomous executor; engineering moves from personal experience to engineering standards to automation to platformization; and the human role shifts from writing code to commanding AI to managing tasks to defining goals and controlling outcomes.

Part 2: Productivity Reconstruction – From Efficiency to Throughput

AI-driven R&D delivers simultaneous improvements in speed, quality, and throughput. Harness integrates AI into the R&D workflow, turning individual coding efficiency into system-wide production efficiency.

Speed (提速): AI participates in requirements analysis, task decomposition, coding, testing, fixing, and delivery, compressing the end-to-end cycle from requirement to delivery.

Quality (提质): Quality control shifts from post-development manual checks to continuous verification embedded in the process. AI combines automated testing, code review, regression, and verification to detect and fix issues.

Throughput (提产): The metric changes from individual lines of code per day to the number of concurrent workflow loops closed. One person can orchestrate multiple agents working in parallel across tasks, repositories, and workspaces.

A data observation from the author's personal code contributions over three years shows total code lines growing from 3k to 20k, while AI-generated code proportion rises from 0% to 90%. The article cautions that line growth does not equal quality or business value growth; AI has become a code productivity amplifier, and the next measure is how much new code translates into effective software delivery.

Summary: Previously, IDEs optimized coding, CI optimized builds, and automated testing optimized verification – each a local optimization. AI connects these local optimizations into a complete production chain. The measure is no longer how much faster each stage is, but how many effective workflows the R&D system can run simultaneously.

Part 3: Organization Reshaping – Four Structural Shifts

AI redefines functional division rather than eliminating roles. Two phenomena illustrate this: Alipay's AFX split (July) moved frontend teams into business lines, shifting direction toward full-stack agent development; ByteDance's QA/RD fusion (ongoing) moves testing from independent quality checks to continuous quality assurance within development.

Four structural shifts:

Functional Boundaries – Individual Scope Expands: Division by professional function (completing one stage) shifts to end-to-end capabilities around business goals (driving one thing to completion). Product, design, development, test, and operations boundaries blur; "full-stack" extends from technology stack to end-to-end work capability.

Execution Abstraction – New Agent Layer: The execution unit moves from services/modules/systems to tasks/workflows/agents. Agents become the new execution unit, connecting requirements, code, tools, and verification systems, with workspace abstraction layered on top.

Role Upgrade – Developer to Builder: The core shifts from code implementation to owning the final outcome: discover problem → define goal → design solution → invoke agents → compose tools → verify result → deliver. Code remains foundational but is no longer the work boundary.

Engineering Capability – Platformization: Reliance on individual tacit experience moves to enterprise-grade infrastructure/platforms (e.g., Harness, internal developer platforms). Repositories, dev environments, CI/CD, testing, permissions, context, agents, observability, and verification become platform capabilities.

Summary: "Foot soldiers" upgrade to "special forces"; individual execution radius expands to the level of a past team or even department.

Part 4: Management Shift – Technical Managers Become System Architects

The paradigm shifts from "personnel division" to "system architecture." Traditional management value (meetings, progress chasing, relaying instructions) is deconstructed by technology; overseer/messenger managers disappear.

Context Tools Eliminate "Messenger" Barriers: Tools like Claude Tag/Projects and GitHub Issue smart linking enable lossless information penetration: requirements, technical specs, and business logic are packaged and solidified. Decentralized alignment lets strategy reach frontline developers and AI directly, without layers of meetings. Information transmission resistance trends to zero; roles surviving on information asymmetry are "pierced."

Traditional Management Functions Deconstructed by AI:

Progress overseer: Analyzes code commits and Jira update frequency to predict delay risks, replacing PMO weekly reports.

Information alignment: Summarizes massive cross-team documents, instantly producing meeting minutes with action items and owners.

Performance evaluation: Outputs objective baselines based on code quality, bug rate, delivery timeliness, eliminating subjective bias.

Management Span and Skill Tree Transfer: "Middle Management Squeeze" – senior leadership penetrates middle layers directly. Shifts: from tracking attendance/daily reports/micromanaging execution → to defining problems and business acumen; from information relay station → to business × engineering "four-dimensional orchestrator"; from rank authority/information asymmetry pressure → to psychological safety and coaching empowerment. Human-AI hybrid teams: directly lead 3 core experts, each with a custom AI agent. Scope shifts from "managing headcount" to "compute orchestration density and human-AI hybrid efficiency."

Designing the "R&D Production System": Build guardrails and validation: architecture standards, security rules, automated review and risk control to prevent high AI output from becoming high risk. The technical manager sits at the center of four pillars: Business Goals (why?), Technical Architecture (how?), AI Capabilities (who does it?), and Organizational Execution (human experts tackle hard problems).

Summary: Managers no longer just assign tasks to teams; they orchestrate business, technology, people, and AI into a more efficient, reliable "R&D production system" – from "task dividers" to "system architects."

Part 5: Frontier Opportunity – Frontend Developers Become FDE Business Problem Solvers

Frontend evolves from "page implementers" to "FDE (Frontline Deployment Engineer) business problem solvers." AI flattens the scarcity of pure UI implementation; frontend upgrades from "implementing requirements" to "defining and solving problems."

Scope Leap – Vertical End-to-End Consumption: Traditional frontend owns a page (UI/interaction) → Full-stack frontend owns a feature (frontend + AI-assisted backend API/database) → FDE era owns a complete business problem (define problem + AI full-chain orchestration + delivery verification). Delivery definition changes: high-fidelity design replication → running data loops and software solutions that solve business problems.

New Role – FDE Frontline Deployment Engineer: Not a "full-stack slave" but an "engineering special forces" deployed to business/customer sites: co-define problems with business, use AI to complete development and integration verification in seconds/days. Single-point development capability ↓ (hand-written components, syntax memorization scarcity goes to zero); Business understanding and abstraction ↑ (hear unspoken real pain points); AI orchestration and system assembly ↑ (mobilize agents, serverless/low-code to quickly assemble MVPs).

Why Frontend Is Naturally Suited for FDE Transition: Human-machine interface empathy: better than backend at interaction/UX, better than product at technical feasibility/boundaries. Instant feedback sensitivity: accustomed to "code change, UI refresh instantly," enabling instant consensus with business during high-frequency MVP alignment.

Bottom-Up Transition Strategy: Upstream: help write PRDs, use Claude Projects to solidify knowledge bases, AI-create 2-3 dynamic models, become requirements co-creator. Downstream: use Cursor/Bolt to close-loop BFF, serverless, basic APIs without waiting for integration scheduling. Choose proving grounds: internal efficiency tools, admin backends, new business MVPs – low risk, high pain, ideal FDE training grounds.

Summary: From "implementing requirements" to "defining and solving problems."

Overall Summary

AI's reshaping of R&D is not a single-point efficiency gain but a systemic reconstruction from coding paradigm to organizational roles: paradigm shifts from humans writing code to humans defining intent and AI delivering; production shifts from individual efficiency to R&D system throughput; organization shifts from functional division to end-to-end builders; management shifts from messengers to system architects; individuals shift from page implementers to FDE problem solvers. AI does not eliminate roles but expands individual execution radius to the scale of a past team – frontline developers upgrade from "foot soldiers" to "special forces."