Amp: The Emperor Has No Clothes

Episode Summary

Executive Summary: SourceGraph's Quinn Slack and Thorsten Ball discuss the evolution of their coding agent from Cody to AMP, emphasizing the need to reset expectations and move fast in a rapidly changing AI landscape. They advocate for building for power users who want the best agent, not mainstream developers, and stress that the model is just one part of a larger system. The conversation covers technical design choices (VS Code vs CLI), the dangers of premature features like MCP and prompt enhancers, and the future of async agents and outer-loop workflows.

Main Topics: Rebranding from Cody to AMP (Priority: 5/5): The decision to create a separate product (AMP) rather than evolve Cody, driven by the need to reset user expectations, avoid disrupting existing contracts, and enable rapid shipping (15x/day). Building for Power Users vs. Mainstream (Priority: 5/5): AMP focuses on developers who want the best coding agent, not the masses. This allows them to ignore features like model choice, bring-your-own-key, and discounting, which slow down innovation. VS Code Extension vs. CLI (Priority: 4/5): Both are supported with a 50/50 internal split. The CLI offers flexibility (SSH, multiple panes, free UI), while VS Code provides better rendering of diagrams and images. The team constantly considers dropping one to reduce complexity. Going with the Grain of the Model (Priority: 4/5): Rather than building elaborate workflows (MCP, custom sub-agents, prompt enhancers), AMP optimizes the harness for each model. The model's capabilities should dictate the product, not the other way around. Async Agents and Outer Loop (Priority: 4/5): The future is 10–100x more agents running concurrently in the background. This creates new challenges in orientation, review, and merge conflicts. The current GitHub-based review model is inadequate. The Changing Role of Engineers (Priority: 3/5): Engineers must now understand product and business, not just write code. The old triangle of PM, designer, engineer is dissolving. Developers need to think about how to wield code, not just produce it. Evals vs. Vibes (Priority: 3/5): AMP deliberately avoids formal evals and instead relies on dogfooding and shipping 20x/day to get real-world feedback. Formal evals are seen as too costly and slow given the pace of change.

Key Arguments: Building a separate brand (AMP) was essential to avoid being slowed down by existing contracts and user expectations from Cody. The market for coding agents is not about supporting multiple models or cheap alternatives; it's about who builds the very best agent for power users. Many popular features (MCP, prompt enhancers, custom sub-agents) are overhyped and actually degrade performance because they don't align with how models are trained. The future of coding agents lies in async agents running 24/7 in the background, not just interactive single-agent use. Engineering teams must rethink processes: no formal code reviews, ship to main, and use the product to build the product (dogfooding). Enterprises should not try to force agents on every developer; instead, they should equip the top 20% who will push the frontier. The value of code written by humans is diminishing; the real value now is in understanding what to build and how to wield agents.

Data Points: AMP growth rate: More than 50% month over month - Growth is often much faster in some weeks, driven by teams of 2-3 people with annual run rates of hundreds of thousands of dollars. Cody team to AMP team size: 8 people on the AMP core team - The team is deliberately small to enable fast shipping (15 times per day) without formal code reviews. Code review adoption on AMP team: 0% formal code reviews - They push to main and rely on dogfooding and fast feedback loops instead of traditional code review processes. Number of models used in AMP: 3+ models from different providers - AMP uses models from Anthropic, OpenAI, and Google, and is close to shipping a fast open-source model as a sub-agent. CLI vs VS Code internal usage: 50/50 split - An internal poll at SourceGraph showed equal preference for CLI and VS Code, with generational differences observed.

Pivotal Quotes: "The only thing that matters is building the best coding agent. Nothing else matters because if you can build that, that's way bigger than anything else that came before." — Quinn Slack: Quinn states the core philosophy behind AMP, prioritizing agent quality over all other features and business concerns. "We alone among the entire industry, it feels like we are being really honest and really bold with that. And I am really concerned just for the rate of progress overall that a lot of these other tools that are great, like Claude Code and Codex and Cursor and so on, that they've forgotten what made them great and what made them grow so fast, which is building the very best product." — Quinn Slack: Quinn critiques competitors for over-fitting on current capabilities and warns that they will peak and decline if they don't stay focused on the core mission. "The model will become an implementation detail to some sense, and we will end up on a different abstraction layer." — Thorsten Ball: Thorsten predicts that specific model names and versions will fade into the background as users interact with agents at a higher level of abstraction.

Implications: The coding agent market is moving toward specialized, high-performance tools for power users. Companies that try to serve everyone with broad features will fall behind. Developers must rethink workflows, tooling, and even codebases for agent consumption. The role of engineers shifts from writing code to directing agents, requiring deeper product and business understanding.