Five Years Ago, I Made a Prediction About Code. Here's My Next One
Over 95% of the author's production code is now generated by AI agents, marking a fundamental shift from manual coding to intention-driven development where developers define requirements and constraints rather than typing syntax The concept of "harness engineering" emerges as the new discipline: software quality depends less on trusting AI models and more on the surrounding system of tests, linters, permissions, and feedback loops that constrain and verify agent output The next generation of Sa
Analysis
TL;DR
- Over 95% of the author's production code is now generated by AI agents, marking a fundamental shift from manual coding to intention-driven development where developers define requirements and constraints rather than typing syntax
- The concept of "harness engineering" emerges as the new discipline: software quality depends less on trusting AI models and more on the surrounding system of tests, linters, permissions, and feedback loops that constrain and verify agent output
- The next generation of SaaS will be "agent-first," meaning services will be designed to be operated directly by AI agents via structured APIs and protocols rather than requiring human interaction with graphical interfaces
- Protocols like Anthropic's Model Context Protocol (MCP) and Google's Agent2Agent (A2A) are foundational infrastructure enabling agents to discover, connect to, and collaborate across services without human GUI navigation
- Software interfaces will transition from being the primary entry point for executing actions to serving as contextual display and decision-approval surfaces, with voice and distributed displays (e.g., AR glasses) becoming new interaction modalities
Why It Matters
This article articulates a paradigm shift that directly affects how AI practitioners, SaaS builders, and software engineers should think about product design and development workflows. The prediction that agent-first SaaS will become the norm means companies that design their services with machine-consumable interfaces, explicit permissions, and structured action schemas will gain a significant competitive advantage. For developers, it signals that the core value is moving upstream from code generation to requirement definition, constraint design, and verification system architecture.
Technical Details
- Agent-first development workflow: The author describes using voice input (SuperWhisper) combined with Codex agents that operate directly within project directories, modifying code based on natural language requirements, constraints, and technical guidance — with human oversight focused on reviewing outputs rather than writing syntax
- Harness engineering: OpenAI's framework for constraining AI agents through project instructions, tool permissions, test suites, linters, and feedback loops; Robert C. Martin's "extreme constraints" approach includes unit tests, Gherkin tests, QA procedures, mutation testing, and coverage metrics as a verification "gauntlet"
- Model Context Protocol (MCP): Released by Anthropic in November 2024, this protocol standardizes how AI applications discover and connect to external data sources and tools, enabling agents to interact with services without navigating graphical interfaces
- Agent2Agent (A2A) Protocol: Created by Google and donated to the Linux Foundation, A2A enables independent AI agents to discover each other's capabilities and collaborate on multi-step tasks across service boundaries
- Distributed interface architecture: The article envisions interfaces appearing contextually within conversations, on multiple devices (phones, work apps, AR glasses like Meta Ray-Ban Display), and triggered by events/webhooks rather than user-initiated navigation
Industry Insight
- SaaS companies should prioritize building machine-readable API contracts, explicit permission models, and well-documented action schemas now — services designed exclusively around graphical interfaces risk becoming obsolete as agent-first consumption becomes standard
- The role of software engineers will increasingly center on "harness engineering": designing robust verification systems, constraint frameworks, and feedback loops rather than writing implementation code, making test architecture and quality gates more valuable than ever
- The separation of product logic from user interface creates an opportunity for new intermediary layers — agents that translate natural language intentions into structured service operations — which could give rise to a new category of AI-native middleware and integration platforms
Disclaimer: The above content is generated by AI and is for reference only.