Our mission is to make Coinbase the best place to build in crypto. To achieve this, we implement a continuous improvement flywheel throughout the company, driven by a strong suite of tools and platforms and a framework that incorporates both quantitative and qualitative metrics. For this post, we will be discussing the AI tools that have been incorporated to help our developers improve the coding journey while maintaining the highest bar for customer safety and quality.
Our approach to consistently improving developer productivity at Coinbase revolves around three key loops within the software development lifecycle. We utilize a combination of quantitative and qualitative measures to assess productivity. Quantitative metrics are built upon the DORA framework, while qualitative insights are gathered through surveys. By establishing a baseline of these metrics, we are able to identify and address areas for improvement.
Our current north star is to grow adoption and increase familiarity with LLM workflows without sacrificing quality, which we believe are necessary steps to empower our engineers. In the spirit of this mission, we have enabled a variety of common coding tools across the company such as Cursor, Copilot, and Claude Code. We’ve also invested in tooling that lets our engineers safely experiment with brand new and custom tools, which can run directly on foundation models through an OpenAI compatible router that is now used daily by >1500 engineers. In addition to agentic and IDE-based tools, we’ve also created a host of internal MCP integrations that allow engineers to create powerful workflows for automation across the SDLC. Here, we share our experience onboarding to Cursor.
We started working with the Cursor team in Summer of 2024 during its meteoric rise in popularity. While various teams had used LLMs for one-off migrations or quality reports, this was the first Coinbase-wide push to fundamentally shift how teams would plan, build, test, and integrate new features using LLMs.
Our DevX team has focused on developing essential MCP servers like Github and Linear integrations that unlocked the true power of Cursor, utilizing standardized patterns for organizing and extending Cursor rules, training engineers on how to leverage different models for varied use cases, and encouraging self-service experimentation
Since then, AI adoption has rapidly increased at Coinbase. By February 2025, every Coinbase engineer had utilized Cursor, which has become the preferred IDE for most of our developers (with some still preferring Cody, JetBrains, or other).
To help drive adoption, we share a set of monthly metrics with eng leaders that includes lead-time-to-change, deployment frequency, bugs, incidents, and AI usage. For lack of a better metric, we initially measured % of code written by AI and are starting to focus on token usage, which we have found to be a good leading indicator of adoption.
These AI metrics are integrated into our broader framework for evaluating software and we anticipate improvements in these metrics as AI tool usage expands across the entire SDLC, leveraging powerful tools like background agents and newer models.
LLMs don't produce flawless code, and we have seen that a growing use of AI in development increases bugs. But AI is still just a tool for us, and tools don’t ship bugs, humans do. To protect against these risks, in close partnership with security and privacy teams, we developed a repository sensitivity matrix that identifies criteria that repository owners must meet to use LLM-based tools to operate on code. This ensures Coinbase maintains the highest bar for customer safety and quality while allowing our engineers to move quickly and with confidence.
One helpful input heuristic we include in our developer metrics is the percentage of AI-generated lines of code relative to human-generated code, and the trend is clear: AI is on track to eclipse human-generated code at Coinbase by the end of the year. This has enabled profound success stories that weren’t possible 12 months ago, like single engineers refactoring, upgrading or building new codebases in days instead of months.
While we have seen a company-wide increase over the last ~3 month period, some parts of the business have increased adoption at faster rates. This is due to the ‘spikey’ nature of an LLM’s usefulness and fit as the right tool for the right job. Leveraging LLMs for coding is not a magic-bullet we should expect teams to universally adopt. Our most senior engineers might spend weeks finding the right fixes to make, and it’s important to focus on results, not any single path to get there.
In looking across the company, we find that teams that adopt LLMs at a faster rate are building frontend UI features, working with less-sensitive data backends, and quickly expanding their unit testing suites. Additionally, rapidly-developed greenfield products benefit greatly from the increased speed of LLM-based workflows. Where we do not see the same meaningful increase in productivity are low-level systems workflows, provisioning and cloud infra teams, complex and system-critical exchange systems, etc. We believe other parts of the development cycle such as planning, test, and observability may benefit these teams more.
This chart represents our initial assessment of the adoption of AI codegen across various orgs within Coinbase, with the dark blue line showing the overall growth across the company. It’s still early days with respect to widespread use across the company partly due to enablement and partly due to strategies to scale growth securely. Also, accurately evaluating AI-driven productivity per developer is nuanced, considering the long-tail impacts of reduced toil and cognitive load on developers. We believe AI will boost productivity tremendously as we tackle bottlenecks in other parts of the SDLC like planning and approval flows.
Our focus today is on enablement and adoption to transform our workforce into "AI-Natives." We are pushing the boundaries further by:
Expanding Tooling: We're deploying both our own and third-party tools to address bottlenecks in code review, testing, and incident resolution.
Embracing Agents: We are starting to experiment with agentic systems that will empower even non-engineering functions to create AI-driven solutions.
Investing in the Future: The investments we make in tools today are preparing us for the models of tomorrow that will be capable of higher contexts and deeper reasoning.
By continuing to invest in our developers and the tools they use, we are not just accelerating our ability to get ideas into production; we are building a culture of innovation to make Coinbase the best place to build.
This journey is just beginning, and we'll continue to share our progress. Stay tuned.
Coinbase has closed our acquisition of Deribit, officially making Coinbase the most comprehensive global crypto derivatives platform. This follows a record month of volume and revenue for Deribit as international momentum around crypto options heats up.
