10xAI Wiki

13分钟阅读

AI Coding Paradigm — From Vibe Coding to Agentic Engineering

The Paradigm Shift

Software development is undergoing its most significant transformation since the introduction of high-level programming languages. The shift is from “coding” to “directing” — developers specify what they want, not how to implement it. Multiple overlapping paradigms describe this shift from different angles.

Vibe Coding

Definition: Using natural language to direct AI to write code, with the developer in a supervisory role. Coined by Andrej Karpathy.

Characteristics:

  • Describe intent in plain language
  • AI generates code
  • Developer reviews and corrects
  • Less emphasis on syntax, more on “vibe” of desired behavior

Key insight: The bottleneck shifts from typing code to understanding what you want. If you can’t describe it clearly, the AI can’t build it well.

Controversy: Peter Steinberger calls “Vibe Coding” a derogatory term. His preference: Agentic Engineering — disciplined, structured use of AI coding agents.

SDD — Specification-Driven Development

Definition: Define specifications first, let AI implement. The spec is the source of truth, not the code.

Workflow (from GLM-5 technical report):

用户需求 → 自然语言规范 → AI生成规范文档 → AI生成代码 → 验证 → 迭代

Key practices:

  • Write specs before code (TDD but for AI)
  • AI checks its own output against the spec
  • Human reviews spec, AI reviews implementation

Why SDD works:

  • Reduces hallucination (code that looks plausible but does wrong thing)
  • Creates a traceable artifact (spec → implementation)
  • Makes code review feasible (human reviews spec, AI-generated code can be spot-checked)

Document-Driven Development

Definition: Shift from “code-centric” to “document-centric” development. Documentation is the primary communication medium between human and AI.

From “AI 原生研发范式：从’代码中心’到’文档驱动’的演进” (2026-02-04):

TraditionalAI-Native
Code is source of truthDocument is source of truth
Documentation is secondaryDocumentation is primary
Humans write codeHumans write specs; AI generates code
Code reviewSpec review + spot-check
Code commentsLiving documentation

Key principle: Write your requirements as if describing to a junior developer who will implement everything — because that developer is an AI.

Agentic Engineering (Peter Steinberger’s Term)

Peter Steinberger’s preferred term for serious AI coding:

“Vibe Coding is a derogatory term. I do Agentic Engineering.”

Characteristics:

  • Disciplined use of AI coding agents
  • Short, precise prompts (not rambling descriptions)
  • Codebase optimized for AI comprehension (clear structure, good naming)
  • Harness Engineering practices (constraints, feedback loops, CI)
  • Testing and verification are not optional

Harness Engineering as the Discipline Layer

The most sophisticated organizations layer Harness Engineering on top of Vibe Coding:

  • CLAUDE.md / AGENTS.md: Project conventions and rules
  • Linters and formatters: Mechanical enforcement of style
  • CI pipelines: Automated testing and verification
  • Sub-agents: Task decomposition and parallelization
  • Feedback loops: AI evaluates its own output

See: Harness Engineering

Tool Landscape (2025-2026)

ToolTypeKey DifferentiatorCompany
Claude CodeCLI AgentDual-agent, hooks, best for complex tasksAnthropic
CursorIDE + AgentGUI-based, 8-agent parallel, Composer modelAnysphere
CodexChatGPT AgentSlack integration, team collaborationOpenAI
QoderCLI + IDEMemory-aware, Chinese market focusQoder Team
WindsurfIDE AgentGood for long tasks, no Claude 4Codeium
DevinWeb AgentGitHub integrated, autonomousCognition
AMPCLI AgentTask completion focusIndependent

Key Insights from Practitioner Reports

Cursor 2.0 (October 2025)

  • Self-developed Composer model (RL-trained MoE)
  • 8 parallel agents in isolated workspaces
  • Speed focus: 4x faster than equivalent models
  • Built-in browser + voice mode

Claude Code Economics (Kieran Klaassen/Every)

  • $250/day cost for intensive usage
  • 2 engineers = 15-person team output
  • “Compounding Engineering” concept: each task makes next tasks easier
  • Parallel agent execution for 6-7 simultaneous tasks

YC Vibe Coding Guide

  • Strict Git usage essential (don’t rely on AI undo)
  • Prioritize end-to-end tests over unit tests
  • Create instruction files (cursor.rules, claude.md)
  • Use multiple tools: Cursor for frontend, Windsurf for long tasks

Agentic Coding Expands Beyond Code — Creative and Media Tools

AI coding agents are expanding beyond pure software development into creative tools:

Codex + HyperFrames: Eating the Video Editing Industry

OpenAI’s Codex combined with HyperFrames is automating video editing workflows, marking a crossover from coding agent to creative production tool. See ../sources/2026-05-06-Codex + HyperFrames 正在吃掉剪辑行业.

Skills as Code: Prompts-as-Products

The ai-translate project shows that two prompt files can be a complete product — installed as Skills in Claude Code / Codex / Cursor. See ../sources/2026-05-05-我开源了一个 AI 轻量翻译工具.

PromptPilot: Structured Prompt Optimization

PromptPilot transforms vague instructions into structured step-by-step directives. See ../sources/2025-06-20-PromptPilot工具介绍与演示.

The Evolution Timeline

YearParadigmKey Activity
2022-2023Copilot (autocomplete)AI suggests next line
2024Chat + Code ReviewAI explains code, reviews PRs
2025Vibe CodingAI writes whole files/features
2025Agentic CodingAI operates autonomously in codebase
2026Harness EngineeringFull discipline layer around AI coding

From Individual to Team: Harness Governance + SDD

The most advanced practice emerging in 2026 combines Harness Engineering with SDD (Specification-Driven Development) at the team level. Key insight from 高德 (Amap) team: 80-90% AI code generation rate does NOT equal real productivity gains — because coding is only ~30% of the full R&D chain (requirements → review → design → development → testing → integration → deployment).

The Three Core Problems

  1. AI-generated code looks correct but has subtle business errors — syntax OK, semantics broken
  2. Legacy codebase risk: Vibe Coding in existing large codebases (100K+ LOC, multiple middleware) is dangerous — implicit dependencies and tacit knowledge are not machine-readable
  3. Single conversation limits: Complex multi-module requirements exceed single AI conversation capacity

The Solution Architecture

  • SDD: Spec as the single source of truth — structured, machine-executable “intent code” replaces prose PRDs. Workflow: Specify → Plan → Implement → Validate
  • Harness Governance: 4 pillars (Context Architecture, Agent Specialization, Persistent Memory, Structured Execution), implemented via .harness/ directory with Rules, Skills, Wiki, and Change Management
  • Qoder Knowledge Base: 3-layer structure (Project Layer / Technology Layer / Asset Layer) with README.md as index for progressive disclosure
  • HITL (Human-In-The-Loop): Developer shifts from coder to intent clarifier + spec reviewer + result validator

See: Harness Engineering, ../sources/2026-05-07-告别氛围编程-Harness治理-SDD-团队级AI研发范式

Erik Schluntz: The Real Vibe Coding

Erik Schluntz (Anthropic researcher) clarifies the true definition:

“As long as you’re reviewing AI-generated code line by line, you’re not vibe coding — you just swapped to a more expensive IDE.”

Verification abstraction layer: Run tests instead of reading code → Experience product instead of running tests → Analyze user data instead of experiencing personally.

AI capability doubling: Independent task duration doubles every 7 months. When AI produces a week’s worth of code in one shot, line-by-line review becomes impossible. See: Erik Schluntz

Boris Cherny’s 2026 Verdict: Programming is Solved

From Boris-Cherny’s Sequoia AI Ascent 2026 keynote:

Key claims:

  • 2026 to date, he has not written a single line of code by hand
  • He no longer uses a computer — all work via Claude App on his phone
  • He manages 5-10 active sessions with hundreds of sub-agents running in parallel
  • Personal record: 150 PRs in a single day
  • Claude Code’s own codebase may shrink to ~100 lines within a year (“code autophagy”)
  • “For a model, everything is just tokens” — developers no longer need to care about local environment, compiler, or language syntax

The Loop paradigm:

  • /loop command makes AI self-loop like a cron job
  • Example loops: auto-fix CI errors, monitor X feedback every 30 min, patch flaky tests
  • Routines (server-side loop) just launched — agents continue working even with laptop closed

Historical analogy: Compares AI revolution to the 15th-century printing press. Pre-press: 10% European literacy. 50 years post-press: more works published than previous 1000 years combined. Book costs dropped 100x. Software will be democratized to the level of texting.

Business moat implications: AI weakens switching costs (easy code migration) and process power (continuous iteration breaks any workflow). Network effects, scale economies, and resource exclusivity remain strong.

Source: 2026-05-07-Claude-Code-father-Sequoia-speech

Sources: Vibe Coding Paradigm, Document-Driven Development, GLM-5: From Vibe to Agentic Engineering, Vibe Coding + SDD

2026-05 Update: From Vibe Coding to Managed Agent Teams

The 2026-05 bulk ingest adds a clear pattern: AI coding is no longer only an individual productivity trick; it is becoming a managed production system.

Working distinction: Vibe Coding is an interaction style; Agentic Engineering is an operating model. The latter requires specs, versioned context, test harnesses, permissions, observability, and review loops.

Architect’s Survival Guide in AI Era (Dennis Doomen, 2026-05)

From Architect Survival Guide:

  • “If you stop writing code, you gradually lose the experience needed to make good architectural decisions”
  • Built a .NET HTTP Mock library with Copilot: gave it GitHub issues, it generated a high-quality open-source project
  • Tests are the “safety net” — reliable tests matter more than code quality when AI generates implementation
  • AI struggles with architectural-level decisions (abstraction design, dependency inversion) requiring holistic judgment
  • Good branch management and code review are essential — AI can silently overwrite human changes
  • Commit messages should record why decisions were made, not just technical details

Agent-Era Productivity Paradox (2026-05)

From Agent Productivity Paradox (Alibaba Aone engineer):

The 1920s electrification paradox: factories replaced steam with electric motors but saw no productivity gains for 30 years until workflow was reimagined. Same pattern in AI era: organizations equip everyone with AI tools but keep industrial-era collaboration structures.

Proposed solutions: All-in-Code monorepo, version everything, agent-safe execution (sandboxes, tiered permissions, dry-run), Agent Teams platform, ChangeSet for unified change tracking, Agentic IAM.

关系图谱

反向链接