The era of the Small Giant (Interview)

What It Covers

Damien Tanner, founder of Pusher and now building LayerCode, returns after seventeen years to discuss how AI coding agents fundamentally reshape software development. The conversation explores why traditional SaaS models face extinction, how code review becomes a bottleneck in the AI era, and why small teams can now build giant companies using tools like Claude Code and Cloudflare Workers.

Key Questions Answered

• •Custom CRM in Minutes: Tanner built a complete CRM system by voice-dictating requirements for fifteen minutes to Claude Code, replacing expensive commercial tools. The system included gamification features tailored to his preferences and could be modified instantly during use. This demonstrates how developers can create bespoke software faster than learning existing SaaS platforms, fundamentally challenging the subscription software model that dominated the past two decades.
• •Post-SaaS Architecture: SaaS interfaces exist for slow humans to navigate complex UIs, but AI agents don't need visual interfaces to complete work. When agents can directly access APIs and databases to accomplish tasks like lead enrichment or data analysis, the traditional web application layer becomes unnecessary. This shift means software moves from human-operated tools to agent-executed workflows, with humans only needing dashboards to review completed work rather than interfaces to perform it.
• •Code Review Elimination: Teams using AI coding agents produce so many pull requests that code review becomes the primary bottleneck, while new teams without review processes move multiple times faster. As language models surpass human code quality and security practices, the entire apparatus of human code review—built to prevent human mistakes—loses relevance. Developers must shift mindset from reviewing code to validating functionality like product managers do.
• •Real-Time Voice Infrastructure: LayerCode provides voice API infrastructure solving complex problems in conversational AI: detecting when users finish speaking despite pauses and filler words, handling interruptions while agents speak, and maintaining sub-one-second response latency. The system runs on Cloudflare Workers across 330 global locations, streams partial transcripts during user speech, and uses Gemini Flash for interrupt detection with 250-300 millisecond response times to determine genuine interruptions versus acknowledgments.
• •Time-to-First-Token Priority: Voice agent quality depends entirely on time-to-first-token latency, not total token throughput. Only Google Gemini and OpenAI GPT-4o optimize for this metric, with most LLMs prioritizing intelligence or throughput instead. Inconsistent latency creates worse user experience than consistent slower responses—when an agent responds in one second initially but takes three seconds on subsequent turns, users assume the system broke rather than accepting variable performance.
• •Plugin Architecture Over RxJS: Tanner rebuilt LayerCode's core from RxJS stream processing to simple plugin architecture with async iterables because the original codebase was incomprehensible to both humans and LLMs. The new waterfall message-passing system allows unit testing individual plugins, enables test-driven development with coding agents, and lets agents successfully add features. This architectural choice prioritized AI agent comprehension over traditional engineering patterns, recognizing that unmaintainable-to-AI code slows development velocity.
• •Ralph Wiggum Development Loop: The autonomous coding pattern involves writing ambitious specifications in a markdown file, then running a shell script that repeatedly calls Claude Code until it returns "complete" in XML tags. Developers leave for walks or lunch while the agent works through entire feature lists. This YOLO mode approach works for greenfield projects where code quality matters less than functionality, and developers can validate outputs without reviewing implementation details.