Sep 28, 2025

September 28, 2025, 5:51 pm IDT In the frenetic world of artificial intelligence, the only constant is change, and the only path to survival is often self-disruption. This profound reality underscored a recent conversation on the Latent Space podcast, where host Alessio Fanelli sat down with Quinn Slack, CEO of SourceGraph, and Thorsten Ball, Amp’s “Dictator” and lead engineer. Their discussion illuminated SourceGraph’s strategic pivot from its established AI coding assistant, Cody, to its latest offering, Amp Code, delving into the philosophy behind building coding agents in a hyper-evolving landscape, the challenges of rapid iteration, and the surprising realities of developer adoption. Thorsten Ball articulated a fundamental realization that catalyzed their shift: “We realized we gotta handle this differently, we gotta reset expectations… you create the new thing that kind of disrupts the business on its own.” This isn’t merely about feature upgrades; it’s about acknowledging that the very definition of a “good” AI coding tool is ephemeral. Quinn Slack underscored this, stating, “The only thing that matters is building the best coding agent. Nothing else matters.” This stark philosophy drives their product development, recognizing that tools relevant just six or twelve months ago are already being superseded. The team consciously chose to “burn the bridge” on some of Cody’s established use cases and customer contracts to fully embrace a new, more agile paradigm with Amp. This radical approach extends to SourceGraph’s internal culture. Thorsten revealed that the Amp core team operates with minimal bureaucracy: “We don’t do formal code reviews. We still push to main. We still ship 15 times every day.” This “founder mode” mentality, unburdened by the slower cycles of larger enterprise software, allows them to dogfood extensively and maintain an unparalleled pace of innovation, directly responding to the rapid advancements in underlying AI models. This internal agility is critical when the external landscape shifts quarterly, not annually. Amp’s market strategy also reflects this adaptability. Instead of aiming for universal enterprise adoption, as was the case with Cody, they selectively target early adopters. Quinn explained, “We pick off the people that want to move as fast as we want to move.” This focus on the “model product frontier” allows them to iterate with users equally invested in pushing boundaries. Interestingly, their internal polling showed a 50/50 split between VS Code extension users and CLI users for Amp, a surprising outcome that led them to consider discontinuing the VS Code extension to further streamline development and responsiveness. A core insight from Ball is that the underlying large language models (LLMs) are rapidly becoming a mere “implementation detail.” While models like Claude 3.5, Claude Code, and Codex CLI dominate headlines today, their individual supremacy is fleeting. “The models will become an implementation detail… the specific model or its version will not be as visible anymore.” The true value, they argue, lies in the “harness” or “scaffolding” built around these models, enabling seamless integration and dynamic adaptation to new, more capable LLMs as they emerge. The relentless pace of AI innovation demands an equally agile development philosophy. Sticking to old paradigms is a recipe for obsolescence. This constant evolution also highlights the evolving nature of AI agent failure modes. Thorsten warned against the “hands-off the wheel” mentality, where developers might blindly trust an agent. He noted, “You can use it in the wrong way and it looks like you’re getting results.” The non-deterministic nature of LLMs means that even a seemingly correct output might be fundamentally flawed, necessitating human oversight and understanding of the underlying architecture. Over-reliance on agents without critical evaluation can lead to “hangover” effects and “spaghetti code.” The market itself is still learning how to effectively leverage these tools, with many users still unsure how to ask the right questions or verify the output. The implications for established software development practices are profound. Traditional roadmaps, long-term architectural planning, and extensive code review processes become liabilities in an environment where fundamental capabilities change every few months. SourceGraph’s success with Amp, shipping 15 times a day with no formal code reviews on the core team, demonstrates a lean, experimental approach that prioritizes learning and speed over rigid adherence to past best practices. This also creates a unique dynamic where the “best” tool is less about a single, perfected product and more about the ability to continuously adapt and integrate the latest advancements. SourceGraph’s pivot to Amp Code embodies a crucial lesson for the tech industry: in the age of generative AI, clinging to established products and development cycles is a direct path to irrelevance. The future belongs to those willing to disrupt themselves, embrace radical agility, and continuously re-evaluate their fundamental assumptions about product, market, and process.