Our Code is Now AI-Generated: A New Framework for the Post-Human Development Era

Listen to the article on Apple Podcasts
Listen to the article on Soundcloud

Abstract

This paper presents the RDP (Research, Development, Production) model as a taxonomic framework designed explicitly for categorizing work in Vibe and Context Engineering — the foundational methodologies of modern agentic AI development. Traditional software engineering approaches have become obsolete in the face of AI agents that can generate, modify, and deploy code autonomously. The RDP model employs a dual-axis classification system using capitalization to distinguish between systemic (capital letters) and component-level (lowercase letters) activities across three primary stages of agentic AI advancement. We demonstrate how Vibe Engineering — structured AI-assisted development where agents are first-class team members — and Context Engineering — the systematic optimization of information flow to AI systems — map to different RDP categories. Through integration with emerging frameworks like the Model Context Protocol (MCP) and analysis of current industry practices where 95% of codebases are AI-generated, we present a framework that captures the rapid, experimental nature of modern AI development. The paper concludes with recommendations for organizational adoption and future extensions as agentic capabilities continue to evolve exponentially.

Introduction

The emergence of Vibe and Context Engineering has rendered traditional software development methodologies obsolete. In February 2025, when Andrej Karpathy coined the term “vibe coding” and described it as “fully giving in to the vibes, embracing exponentials, and forgetting that the code even exists” (Karpathy, 2025), he articulated a fundamental shift already underway in the industry. With 25% of Y Combinator’s Winter 2025 batch reporting codebases that are 95% AI-generated (Fortune, 2025), and teams of 10 engineers now accomplishing what previously required 100 (Tan, 2025), we have entered an era where human-written code is the exception rather than the rule.

Context Engineering, defined as “the delicate art and science of filling the context window with just the right information for the next step” (Karpathy, 2025), provides the systematic foundation that makes Vibe Engineering possible at scale. Together, these methodologies represent not an evolution of traditional development but a complete paradigm shift. The waterfall models, agile sprints, and even DevOps pipelines of the previous era assume human-centric development with AI as an assistant. In contrast, Vibe and Context Engineering position AI agents as primary developers with humans providing strategic direction and context optimization.

The RDP model presented in this paper provides a taxonomic framework specifically designed for this new reality. Unlike traditional R&D classifications that assume slow, deliberate progress through stages, the RDP model captures the rapid, experimental, and often non-linear nature of AI-driven development. By distinguishing between systemic and component-level work across Research, Development, and Production phases, the framework enables organizations to categorize and manage the diverse activities that comprise modern agentic AI programs.

Theoretical Framework

The RDP model for Vibe and Context Engineering operates on fundamentally different principles than traditional development taxonomies. Where conventional frameworks assume sequential progression and human-controlled processes, the RDP model recognizes that AI agents can simultaneously research, develop, and deploy solutions in timeframes measured in hours rather than months. The theoretical foundation rests on three core principles that distinguish agentic AI development from its outdated predecessors.

First, the principle of agent autonomy acknowledges that modern AI systems are not tools but collaborators. In Vibe Engineering, agents make architectural decisions, implement complex features, and even refactor entire codebases without human intervention (Wang, 2025). This autonomy requires rethinking how we categorize work — the distinction is no longer between basic and applied research but between work that advances agent capabilities systemically versus work that implements specific agent functions.

Second, the principle of context primacy recognizes that the quality and structure of information provided to agents determines outcomes more than algorithms or computing power. Context Engineering has evolved from simple prompt engineering to encompass memory hierarchies, tool orchestration, retrieval-augmented generation, and multi-agent coordination protocols (LangChain, 2025). The RDP model must capture activities ranging from fundamental research on context compression to production deployment of context management systems.

Third, the principle of exponential velocity reflects the reality that development cycles have compressed from months to hours. The concept of “bolts” replacing sprints in the AI-Driven Development Lifecycle (AI-DLC) methodology captures this acceleration (AWS, 2025). Traditional taxonomies, which assume careful progression through readiness levels, become meaningless when an AI agent can conceive, implement, test, and deploy a feature within a single working session.

The Enhanced RDP Framework for Agentic AI

The RDP model employs capitalization to distinguish scope and impact within each phase, creating six primary categories tailored specifically for Vibe and Context Engineering activities. This notation system captures both the what (Research, Development, Production) and the how (systemic versus component-level) of modern agentic AI work.

Research (R/r) in the context of Vibe and Context Engineering encompasses activities that advance our understanding of how to build and optimize agentic systems. Big Research (R) includes developing new agent architectures, creating novel context management frameworks, establishing theoretical foundations for multi-agent coordination, and pioneering memory systems that enable long-term agent learning. Examples include Anthropic’s work on Constitutional AI that shapes agent behavior systemically (Anthropic, 2025) or research into attention mechanisms that will allow million-token context windows.

Little Research (r) focuses on specific techniques and optimizations within established frameworks. This includes investigating optimal prompting strategies for particular tasks, testing context compression algorithms, validating agent behavior in controlled scenarios, and exploring tool integration patterns. A team studying how to structure React component requests for maximum agent accuracy would be conducting ‘r’ work — valuable but localized in impact.

Development (D/d) represents the transformation of research insights into functional agentic systems. Big Development (D) encompasses creating comprehensive agent frameworks like LangChain or CrewAI, building production-grade context orchestration platforms, implementing multi-agent coordination systems, and establishing the infrastructure that enables other teams to deploy agents. When Windsurf created their integrated memory system for code generation (TechCrunch, 2025), they were conducting ‘D’ work that would enable thousands of downstream applications.

Little Development (d) includes creating specific agent tools or capabilities, implementing individual memory modules or context processors, building connectors for particular data sources, and optimizing existing components for performance. A developer creating a custom tool for their agent to access internal APIs would be performing ‘d’ work — necessary but not transformative to the field.

Production (P/p) covers the deployment and operation of agentic systems in real-world environments. Big Production (P) involves deploying platform-level agent systems that serve multiple use cases, establishing production context management infrastructure, creating agent orchestration platforms that others build upon, and maintaining the systems that enable an organization’s entire agentic AI strategy, when Microsoft deploys Copilot across its whole Office suite, that represents ‘P’ level work.

Little Production (p) encompasses deploying specific agents for defined tasks, tuning context windows for particular workflows, implementing monitoring for individual agent behaviors, and maintaining focused agent applications. A company deploying a customer service agent for their website engages in ‘p’ work — valuable to their business but not ecosystem-defining.

Application to Vibe Engineering

Vibe Engineering activities map naturally to the RDP framework, with distinct patterns emerging at each level. Research in Vibe Engineering (R) involves fundamental investigations into how AI agents can best collaborate with humans and each other. This includes developing new paradigms for agent-human interaction, creating frameworks for agent specialization and delegation, establishing principles for verification-driven development where agents generate their tests, and researching emergent behaviors in multi-agent coding systems.

Current examples of ‘R’ level Vibe Engineering include studies on optimal team compositions mixing human and AI developers, research into “memory bank” patterns that maintain project context across sessions, and investigations of how agents can learn from codebases to match organizational styles. The shift from individual productivity to team-AI hybrid workflows represents systemic research that reshapes entire development paradigms.

Component-level Vibe Engineering research (r) focuses on specific techniques for improving agent coding performance. This includes A/B testing different prompting strategies for code generation, measuring the impact of various context window configurations, optimizing agent performance for specific programming languages or frameworks, and studying how formatting affects agent comprehension. While less transformative than systemic research, these focused investigations accumulate into significant productivity gains.

Development activities in Vibe Engineering show clear stratification between framework and tool creation. Big Development (D) efforts include building integrated development environments designed for agent-first workflows (like Cursor’s Agent Mode), creating verification frameworks that validate agent-generated code automatically, establishing project template systems that embed organizational knowledge, and developing the Plan-and-Act architectures that enable agents to design before implementing.

The progression to Production reveals the current state of Vibe Engineering adoption. Big Production (P) deployments remain rare but impactful — organizations like Domu Technology that generate 95% of their code through AI represent the vanguard. These deployments require comprehensive monitoring systems for agent-generated code quality, governance frameworks ensuring compliance and security, rollback mechanisms for agent errors, and continuous learning systems that improve agent performance over time.

Application to Context Engineering

Context Engineering activities within the RDP framework reveal different patterns than Vibe Engineering, reflecting its nature as the foundational layer enabling agent capabilities. Big Research (R) in Context Engineering tackles fundamental challenges like breaking the context window limitations that constrain current systems, developing new architectures for long-term memory and retrieval, creating theoretical frameworks for multi-modal context integration, and establishing principles for context inheritance in hierarchical agent systems.

The Model Context Protocol (MCP) developed by Anthropic exemplifies ‘R’ level Context Engineering — it establishes universal standards for how AI assistants connect to external data sources, fundamentally changing how context flows through agentic systems (Anthropic, 2024). Similarly, research into attention mechanisms that maintain accuracy across million-token contexts represents systemic advances that enable entirely new categories of applications.

Little Research (r) in Context Engineering includes focused studies on retrieval-augmented generation (RAG) optimization, prompt template effectiveness for specific domains, context compression techniques for cost reduction, and memory selection algorithms for relevance. While these investigations may seem narrow, they often yield immediate practical benefits — a 10% improvement in context compression can dramatically reduce operational costs at scale.

Development in Context Engineering is experiencing explosive growth as organizations rapidly implement research insights. Big Development (D) encompasses several key initiatives, including the development of production-grade RAG systems such as LlamaIndex, the creation of orchestration frameworks like LangGraph for managing complex context flows, the implementation of distributed memory systems for enterprise-scale deployments, and the establishment of context namespacing systems to facilitate multi-team coordination. These platforms become the infrastructure upon which thousands of agentic applications are built.

Production deployment of Context Engineering systems reveals both successes and challenges. Big Production (P) implementations, such as enterprise-wide knowledge bases accessible to all agents, production context orchestration handling millions of daily requests, and multi-tenant memory systems serving diverse agent populations, demonstrate the technology’s maturity. However, organizations report that context management often becomes the bottleneck — Anthropic’s research system consumes 15x more tokens than single-agent alternatives due to context overhead (Anthropic, 2025).

Integration with Modern Development Practices

The obsolescence of traditional development practices necessitates new organizational structures and workflows aligned with Vibe and Context Engineering principles. The concept of “sprints” becomes meaningless when agents can implement features in hours; instead, organizations adopt “bolts” — rapid development cycles where humans define objectives and agents execute implementation (Salesforce DevOps, 2024). Daily standups evolve into context synchronization sessions, ensuring teams provide agents with access to relevant information.

Code review processes undergo a fundamental transformation in agentic development. Traditional line-by-line reviews become impossible when dealing with thousands of lines of AI-generated code; instead, teams focus on architectural reviews, test coverage validation, and behavior verification. The role of senior engineers shifts from code quality gatekeepers to context architects who ensure agents receive appropriate information for their tasks.

Version control systems adapt to track not just code changes but context evolution. Organizations implement “context versioning” where prompt templates, memory configurations, and tool definitions are tracked alongside traditional code. This enables teams to understand not just what changed but why the agent made specific decisions — crucial for debugging and improvement.

The emergence of “viability engineers” represents a new role combining traditional QA with prompt engineering and context optimization. These specialists ensure that agent-generated code meets not just functional requirements but broader organizational goals around maintainability, security, and performance. They develop test suites that validate agent behavior across various contexts, ensuring consistent output quality.

Case Studies in RDP Classification

Real-world applications of the RDP model illuminate its practical utility. Consider a financial services firm implementing automated trading strategies through agentic AI. Their journey begins with ‘R’ level work: researching how agents can understand market dynamics and developing novel architectures for real-time decision making. This progresses to ‘r’ studies on optimal context windows for price data and backtesting various prompt strategies.

Development phases show clear RDP stratification. The firm’s ‘D’ work involves building a comprehensive trading agent framework with risk management, position sizing, and market analysis capabilities. Simultaneously, ‘d’ efforts create specific tools for accessing market data feeds, implementing order execution protocols, and generating compliance reports. The distinction helps management allocate resources appropriately — ‘D’ work receives long-term investment while ‘d’ work operates on shorter cycles.

Production deployment reveals the framework’s value for risk assessment. ‘P’ level work establishing the production trading infrastructure requires extensive testing, regulatory compliance, and failsafe mechanisms. In contrast, ‘p’ deployments of specific trading strategies can proceed more rapidly with appropriate safeguards. The capitalization immediately communicates the scope and risk level to all stakeholders.

Another illustration comes from a healthcare organization implementing diagnostic agents. Their ‘R’ research into medical reasoning and context integration differs qualitatively from ‘r’ studies on formatting patient data for agent consumption. The RDP model helps them recognize that while both activities are valuable, they require different expertise, timelines, and success metrics. This clarity enables better project planning and resource allocation.

Challenges and Limitations

The rapid evolution of Vibe and Context Engineering creates classification challenges for the RDP model. The boundary between Research and Development blurs when agents can implement novel ideas immediately — does an agent discovering a new algorithm while coding represent ‘R’ or ‘D’ work? Similarly, the distinction between Development and Production becomes fluid when agents deploy their own code automatically.

Context window limitations create practical constraints on system complexity. Despite theoretical advances, production systems struggle with context management at scale — each additional piece of context increases latency and cost. Organizations report that context engineering often becomes the bottleneck, with agents performing well in isolation but struggling when integrated into larger systems requiring extensive context.

The “black box” nature of agent decision-making complicates classification and governance. When an agent refactors an entire codebase overnight, determining whether this represents ‘d’ optimization or ‘D’ architectural change requires understanding the agent’s reasoning — often opaque even to its creators. This ambiguity challenges traditional project management and accountability structures.

Economic considerations limit widespread adoption of advanced RDP categories. While ‘R’ and ‘D’ work in Context Engineering promise dramatic improvements, the computational costs remain prohibitive for many organizations. Multi-agent systems consuming 15x more resources than single agents restrict ‘P’ deployments to high-value use cases where performance gains justify expenses.

Future Directions

The RDP model must evolve with the exponentially advancing capabilities of agentic AI. Near-term developments (2025–2026) will likely see the emergence of self-modifying agents that blur current category boundaries. When agents can enhance their architectures, the distinction between Research and Development becomes philosophical rather than practical. The model may need additional notation to capture these reflexive improvements.

Quantum computing integration presents another evolution vector. As quantum-classical hybrid systems become viable, new RDP categories may emerge for quantum-enhanced agents (‘Rq’, ‘Dq’, ‘Pq’). These systems promise to solve context optimization problems that currently constrain agent capabilities, potentially enabling true autonomous research agents that advance the field without human direction.

The convergence of Vibe and Context Engineering suggests future methodologies that transcend current distinctions. “Contextual Vibing” might describe agents that dynamically adjust their development approach based on available context, while “Vibe Synthesis” could capture agents that learn development patterns from observing human-AI collaboration. The RDP model must remain flexible to accommodate these emergent practices.

Regulatory frameworks will inevitably shape RDP evolution. As governments grapple with AI-generated code’s implications for liability, security, and intellectual property, new categories may emerge for compliance-focused activities. ‘Rc’ might denote research into verifiable agent behaviors, while ‘Pc’ could represent production deployments meeting specific regulatory standards.

Implications for Organizations

Organizations adopting the RDP model for Vibe and Context Engineering must fundamentally restructure their approach to technology development. Traditional IT departments, which are organized around languages, platforms, or products, are becoming obsolete. Instead, teams are aligning around RDP categories, with specialists in systemic research, component development, or production deployment of agentic systems.

Performance metrics require a complete overhaul. Lines of code, commit frequency, and bug counts become meaningless when agents generate thousands of lines hourly. Instead, organizations must measure context efficiency, agent autonomy levels, and business outcome velocity. A developer’s value shifts from personal productivity to their ability to orchestrate multiple agents effectively.

Investment strategies must account for the exponential nature of agentic AI advancement. Traditional ROI calculations, assuming linear improvement, fail when agent capabilities double every few months. Organizations must adopt portfolio approaches, investing across RDP categories with the understanding that breakthrough advances in ‘R’ can obsolete entire ‘P’ deployments overnight.

Cultural transformation proves the most significant challenge. Engineers accustomed to crafting code must embrace roles as agent orchestrators. Architects who prided themselves on elegant designs must accept that agents may produce superior solutions through brute-force exploration. The humility to recognize human limitations while maintaining strategic control requires a psychological adjustment that many find difficult.

Conclusion

The RDP model provides a crucial framework for navigating the complex creativity of Vibe and Context Engineering. By distinguishing between systemic and component-level work across Research, Development, and Production phases, organizations can categorize and manage the diverse activities that comprise modern agentic AI development. The framework’s elegance lies not in rigid categorization but in providing a shared vocabulary for discussing work that often defies traditional classification.

As we stand at the threshold of truly autonomous development, where agents not only write code but also architect systems, manage deployments, and even improve themselves, frameworks like RDP become critical infrastructure. They enable us to maintain strategic coherence while embracing tactical chaos, to govern without constraining, and to advance collectively while allowing individual exploration.

The obsolescence of traditional software development is not a future possibility but a present reality. Organizations that cling to waterfall, agile, or even DevOps methodologies will struggle to compete with those that adopt Vibe and Context Engineering. The RDP model offers a bridge from the old world to the new, providing structure for those ready to embrace development at the speed of thought rather than the speed of typing.

The journey from human-written to AI-generated code represents more than technological evolution — it demands fundamental reconsideration of creativity, authorship, and value creation in the digital age. The RDP model, by providing a taxonomy for this new landscape, enables us to navigate with purpose rather than drift with the current. In a world where code writes itself, our role shifts from creation to curation, from implementation to imagination, from developers to directors of digital symphony. The RDP model ensures we conduct ourselves with clarity.

References

Anthropic. (2024). Introducing the Model Context Protocol. https://www.anthropic.com/news/model-context-protocol

Anthropic. (2025). Building effective AI agents. https://www.anthropic.com/engineering/building-effective-agents

Anthropic. (2025). How we built our multi-agent research system. https://www.anthropic.com/engineering/multi-agent-research-system

AWS. (2025). AI-driven development life cycle: Reimagining software engineering. Amazon Web Services Blog. https://aws.amazon.com/blogs/devops/ai-driven-development-life-cycle/

Fortune. (2025). Silicon Valley CEO says ‘vibe coding’ lets 10 engineers do the work of 100. https://fortune.com/2025/03/26/silicon-valley-ceo-says-vibe-coding-lets-10-engineers-do-the-work-of-100-heres-how-to-use-it/

GitHub. (2025). Context engineering: A practical, first-principles handbook. https://github.com/davidkimai/Context-Engineering

IBM. (2025). What is vibe coding? IBM Think. https://www.ibm.com/think/topics/vibe-coding

Karpathy, A. (2025). There’s a new kind of coding I call “vibe coding” [Twitter post]. https://x.com/karpathy/status/1886192184808149383

LangChain. (2025). Context engineering for agents. https://blog.langchain.com/context-engineering-for-agents

LangChain. (2025). The rise of context engineering. https://blog.langchain.com/the-rise-of-context-engineering

Salesforce DevOps. (2024). The evolution of AI in software engineering: From enhanced TODOs to autonomous virtual employees. https://salesforcedevops.net/index.php/2024/07/01/the-evolution-of-ai-in-software-engineering/

Tan, G. (2025). Personal communication on Y Combinator Winter 2025 batch statistics.

TechCrunch. (2025). Vibe-coding startup Windsurf launches in-house AI models. https://techcrunch.com/2025/05/15/vibe-coding-startup-windsurf-launches-in-house-ai-models/

TRM Labs. (2025). Vibe engineering in the age of vibe coding. https://www.trmlabs.com/resources/blog/vibe-engineering-in-the-age-of-vibe-coding

Wang, F. (2025). The vibe engineering manifesto. https://www.vibeengineering.ai/p/the-vibe-engineering-manifesto

Wikipedia. (2025). Vibe coding. https://en.wikipedia.org/wiki/Vibe_coding