Why did Sourcegraph create a new product (AMP) instead of updating Cody?

The rapid pace of AI model development meant that Cody's existing architecture and brand contracts would hinder rapid iteration. Creating AMP as a distinct product with a new brand allowed Sourcegraph to move much faster and embrace the changes in the AI landscape.

How fast is AMP growing and who are its primary users?

AMP is experiencing rapid growth, exceeding 50% month-over-month. It targets teams and individuals who want to stay at the forefront of AI model capabilities, including small teams with high annual run rates.

What are the core development principles driving AMP?

AMP emphasizes moving extremely fast, shipping frequently (15 times a day), and embracing change. They prioritize a flexible codebase and a mindset that anticipates new technologies rather than relying on established, slower processes.

What is the significance of the VS Code extension vs. CLI support for AMP?

AMP initially focused on a VS Code extension for easier distribution but found that a CLI offers more flexibility for various environments and workflows. They now support both, acknowledging a 50/50 split in user preference and the advantages of each.

Does AMP prioritize certain models or offer model choice?

AMP focuses on building the best coding agent, prioritizing performance and capability over model choice for the user. They believe models will become implementation details, and the focus should be on the system prompt, tools, and scaffolding around the model.

What are the key challenges or failure modes of current coding agents?

Common failures include 'vibe coding' (unproductive use of agents), outsourcing thinking without understanding, and the non-deterministic nature of LLMs leading to unreliable workflows. AMP aims to mitigate these by focusing on core capabilities and user clarity.

How does AMP handle the rapid evolution of AI models and tools?

AMP's strategy embraces constant change, viewing models as rapidly evolving components. They prioritize a flexible 'scaffolding' rather than rigidly integrated features, allowing them to adapt quickly to new models and technologies.

What is Sourcegraph's philosophy on building for early adopters vs. the mainstream?

Sourcegraph intentionally focuses on 'frontier' or early adopter users who are curious and want to push the boundaries of AI agents. They believe this focus is crucial for staying ahead, as building for the mainstream too early would lead to falling behind in the rapidly changing AI landscape.

How is the role of engineers changing with the rise of coding agents?

Engineers increasingly need to understand business and product strategy, not just technical implementation. Coding agents allow engineers to be closer to the customer and product vision, enabling faster iteration from idea to draft.

What are the potential benefits of using multiple models within AMP?

Different models have unique strengths and weaknesses. By integrating models from various providers (Anthropic, OpenAI, Google), AMP can leverage the best suited model for specific tasks or sub-agents, optimizing performance and capabilities.

How does AMP approach team structure and development in a fast-changing environment?

Sourcegraph empowers individuals like Thirstston with significant trust and autonomy ('dictator'). They prioritize first-principles thinking, rapid iteration, and embracing experimentation, even if it involves mistakes, to build the future of coding agents.

Amp: The Emperor Has No Clothes

Latent Space Podcast

Science & Technology8 min read82 min video

Sep 25, 2025|19,009 views|134|10

Save to Pod

Key Moments

On this page

TL;DR

AMP Code redefines coding agents, prioritizing rapid iteration and core functionality over feature bloat, focusing on power users.

Key Insights

Sourcegraph transitioned from Cody to AMP Code to embrace new AI model capabilities and a faster development cycle.

AMP prioritizes building the best coding agent and moving fast, even if it means a new brand and resetting user expectations.

The rapid pace of AI evolution necessitates a flexible development approach, with AMP shipping 15 times daily.

AMP focuses on 'power users' and 'early adopters' rather than trying to cater to the entire market, believing this approach will lead to better product development.

The company leverages existing trust from Sourcegraph to fund 'crazy stuff' with AMP, allowing for rapid experimentation.

AMP differentiates itself by avoiding unnecessary complexity and focusing on core functionality, sometimes even removing features.

The future of coding agents will likely involve more asynchronous operations and a greater focus on natural language interfaces for complex tasks.

FROM CODY TO AMP: A STRATEGIC PIVOT

Sourcegraph's journey from Cody to AMP Code signifies a strategic shift driven by the rapid advancements in AI models, particularly with the advent of Claude 3.5. The team recognized that the capabilities unlocked by newer models demanded a different approach, one that allowed for unconstrained exploration and rapid iteration. This led to the creation of AMP, a new project with a distinct brand and reset expectations for users, separating it from the existing Cody product and its enterprise contracts. This move was essential to avoid being disrupted by the pace of AI innovation.

THE PURSUIT OF THE ULTIMATE CODING AGENT

The core philosophy behind AMP is the singular focus on building the best possible coding agent. The developers acknowledge that no one has truly achieved this yet, but their rapid iteration cycle of shipping 15 times a day allows them to get closer. This relentless pursuit of improvement is contrasted with the market's tendency for tools to have a short lifespan, often becoming irrelevant within 6-12 months. AMP's distinct branding and independent development allow for the agility needed to stay at the forefront of this rapidly evolving landscape.

AGILITY AND CONSTANT ITERATION: AMP'S DEVELOPMENT CYCLE

AMP's technical architecture is designed for extreme agility, enabling a release cycle of 15 times per day. This is a stark contrast to traditional development tied to larger platform release cycles. This rapid iteration is not just about speed but about resetting expectations for how software is built and delivered. By creating a new project with a new way of working, Sourcegraph can adapt more effectively to the swift changes in AI technology, ensuring the product remains at the cutting edge without being encumbered by legacy constraints.

FOCUS ON POWER USERS AND RAPID GROWTH

AMP consciously targets 'power users' and 'early adopters' who are eager to leverage the latest AI capabilities. This strategic decision allows the company to move at a pace that might be too aggressive for a broader enterprise audience. The rapid growth, exceeding 50% month-over-month, validates this approach. By focusing on users who want to stay on the product frontier, AMP can gather crucial feedback and iterate quickly, ensuring it remains relevant in a fast-changing market.

DECOUPLING FROM LEGACY: THE LIBERATION OF AMP

The decision to create AMP as a separate entity from Sourcegraph's established products, like Cody, provided significant liberation. It allowed the team to discard preconceived notions about software development and established processes. This freedom from old habits, like mandatory code reviews or complex planning, enables faster innovation. The ability to 'dogfood' the product extensively, using AMP to build AMP, provides immediate and valuable feedback loops, proving more effective than traditional processes in this dynamic environment.

THE CLI VS. IDE EXTENSION DEBATE AND FUTURE INTERFACES

AMP's dual availability as both a VS Code extension and a CLI reflects a pragmatic approach to user accessibility and technical advantage. While the VS Code extension offers ease of distribution and richer UI capabilities, the CLI provides flexibility, enabling usage across various editors and environments via SSH. The team was surprised by the strong adoption of the CLI, highlighting the evolving ways developers interact with tools. Future interfaces, including potential web and even mobile interactions with async agents, are being considered, emphasizing adaptability as a core principle.

NAVIGATING THE EVER-CHANGING AI LANDSCAPE

The AI tooling landscape is in constant flux, with new models and capabilities emerging rapidly. AMP's strategy is to remain agile and position itself to react quickly to these changes. This means building with a flexible codebase and managing user expectations appropriately. The company acknowledges that what is state-of-the-art today may be outdated in months, underscoring the importance of a development philosophy that embraces continuous change and avoids over-engineering for current capabilities.

THE DELIBERATE AVOIDANCE OF UNNECESSARY COMPLEXITY

A key differentiator for AMP is its deliberate rejection of features that add complexity without proportional value. This includes features like prompt enhancers, custom sub-agents with convoluted token usage, or overly granular MCP tools. The philosophy is to build a lean product that wraps the core model effectively, focusing on what truly drives user value. This disciplined approach allows the team to stay nimble and avoid getting bogged down in maintaining features that might quickly become obsolete or are not aligned with the primary goal of building the best coding agent.

THE DUAL NATURE OF THE CODING AGENT MARKET

AMP acknowledges the existence of different market segments, including those seeking basic, cheaper AI tools. However, their focus remains firmly on users demanding the absolute best coding capabilities, even if it means a higher cost or a more specialized tool. This strategy allows them to push the boundaries of AI development without being constrained by the needs of the mainstream market. They believe this focus on the 'frontier' will ultimately drive broader industry progress.

MODEL CHOICE AND THE ROLE OF THE 'HARNESS'

While AMP utilizes models from various providers (Anthropic, OpenAI, Google), they deliberately do not expose model choice directly to users. The emphasis is on optimizing the 'harness' – the system prompt, tools, and scaffolding around the model – for the best possible performance. The underlying models are seen as an implementation detail, with the focus being on delivering a consistently high-quality experience. This approach ensures that even if a new, faster, or specialized model becomes available, AMP can integrate it seamlessly without altering the user's interaction.

THE EVOLVING DEFINITION OF A 'CODING AGENT'

The conversation around what constitutes a 'coding agent' is evolving. AMP's focus is not on features that might appeal to users of current AI-first tools (like direct chat integrations within an IDE) but on future interaction paradigms. They believe that relying on current models for tasks like in-editor chat might not be the most valuable long-term interaction. The emphasis is on building for the next generation of AI capabilities, which they anticipate will be more asynchronous and integrated in ways not yet fully realized.

ADDRESSING FAILURE MODES AND USER EXPECTATIONS

AMP is highly conscious of the potential failure modes of AI agents, such as 'vibe coding' or outsourcing thinking without understanding. They actively discourage scenarios where users blindly trust agents without proper oversight. The emphasis is on empowering engineers to use agents as powerful tools to augment their own expertise, rather than as replacements for critical thinking. This involves educating users on the limitations and the importance of maintaining control and understanding the outcomes.

THE IMPACT OF ASYNC AGENTS AND THE OUTER LOOP

The rise of asynchronous agents operating in the background presents new challenges for developer workflows, particularly in managing and reviewing their output. AMP is exploring solutions to help users orient themselves with agent progress and understand changes effectively. This includes concepts like unified logging and improved UIs that provide at-a-glance summaries of agent actions. The goal is to create a seamless experience where agents can perform tasks efficiently without overwhelming the developer's ability to track and integrate their work.

RETHINKING VERSION CONTROL AND INTER-AGENT COMMUNICATION

The emergence of multiple, independently operating agents raises questions about traditional version control systems like Git. AMP is considering how to manage potential merge conflicts and ensure efficient cross-agent orchestration. This involves exploring mechanisms for agents to communicate about the files they are touching and to integrate changes effectively. The challenge lies in balancing this coordination with the risk of wasted tokens if changes are frequently rewritten or deemed unnecessary.

ADAPTING CODEBASES AND TOOLING FOR AGENTS

A significant trend is the adaptation of codebases and development tooling to better serve coding agents. This includes improved help text and structured logging that aids agent comprehension. AMP believes that as agents become more powerful, developers might even modify their coding conventions and tooling to optimize for agent interaction. This shift signifies a maturing understanding of how AI can be integrated into the development lifecycle, moving beyond simple code generation to a more symbiotic relationship.

THE FUTURE OF SOFTWARE: USER-GENERATED CONTENT AND POWER USERS

The conversation extends to the notion of user-generated content and the empowerment of 'power users' who can unambiguously specify their needs to a computer. This may blur the lines between technical and non-technical users, as individuals with strong problem-solving skills can leverage AI tools like AMP to build complex solutions without traditional coding expertise. AMP's focus on these users aims to unlock new levels of productivity and innovation, democratizing the creation of sophisticated software.

THE STRATEGIC ADVANTAGE OF TARGETING EARLY ADOPTERS

AMP's strategy of focusing on early adopters and 'frontier' users is a deliberate choice to avoid the pitfalls of trying to cater to the entire market. By concentrating on those who actively seek advanced AI capabilities and are willing to adapt to new workflows, AMP can accelerate its development and maintain a competitive edge. This approach allows them to build a product that truly serves the needs of those pushing the boundaries of what's possible with coding agents.

THE 'BIGGER PICTURE' MINDSET FOR PRODUCT DEVELOPMENT

The development of AMP is guided by a first-principles approach and a long-term vision. The team is less concerned with short-term successes or failures and more focused on building the future of coding agents. This requires a mindset of radical experimentation, embracing mistakes as learning opportunities, and empowering individuals like Thorsten Ball ('the dictator of AMP') to drive innovation. The ultimate goal is to create a powerful coding agent that will shape how software is built for years to come.

Mentioned in This Episode

●Software & Apps

●Companies

Common Questions

AMP is a new AI coding agent from Sourcegraph, built for speed and adaptability, differentiating itself from Cody by resetting expectations and allowing for faster iteration cycles. It was developed to handle the rapid evolution of AI models and agent capabilities.